EP4334741A1 - Method for two-dimensional and three-dimensional imaging based on collocated multiple-input multiple-output radars - Google Patents

Abstract

Method for estimating the angular coordinates of point targets whose range has been estimated through a MIMO FMCW radar equipped with a plurality of transmitting (TX) and receiving (RX) antennas, wherein each couple of said TX and RX antennas is replaced with the equivalent virtual antenna; said virtual antennas form a virtual array composed by one or multiple horizontal uniform linear arrays (HULAs); said MIMO FMCW radar being arranged to generate real or complex signals in response to a propagation scenario including a plurality of point targets; said method including acquisition of a spectrum of said signal and its first derivatives, acquisition of a list of a predetermined number of discretized ranges, named frequency bins, of said previously detected point targets, and sequential execution for each of said frequency bins; said execution including : estimation of the horizontal spatial frequency and complex amplitude of the most dominant point target, the estimation including: computation of the spectrum its first three derivatives through FFT calculation, checking whether combination of the spectra associated with said Reference HULA, generates an amplitude peak in the resulting spectrum and when the checking is positive execution of the following steps: refinement of said range and complex amplitude of said most dominant point target; cancellation of said most dominant point target, calculation of a residual energy in the considered frequency bin and its comparison with a predetermined threshold, and when a residual energy exceeds is below the predetermined threshold computation of the spatial coordinates, otherwise updating of said acquired list of said frequency bins and return analyze a further frequency bin; finally, generation of an overall image of the propagation scenario.

Description

Method for Two-Dimensional and Three-Dimensional Imaging based on Collocated Multiple-Input Multiple-Output Radars

Field of the invention

The present invention relates to a deterministic method for the detection of multiple targets and the estimation of their spatial coordinates in a colocated multiple-input multiple-output radar system operating in a slowly changing two-dimensional or three- dimensional propagation scenario. Two specific embodiments of the proposed method are described in detail for both frequency modulated continuous wave radars (FMCW) and stepped frequency continuous wave (SFCW) radars. These embodiments have, among others, the following relevant features: a) they exploit novel algorithms for estimating the parameters of multiple overlapped sinusoids or complex exponentials affected by additive noise; b) they require the evaluation of one-dimensional fast Fourier transforms only and the search for the maxima of proper cost functions over one-dimensional (frequency, azimuth or elevation) domains; c) they generate radar images in the form of clouds of points.

The second feature make them computationally efficient. Moreover, the first embodiment and part of the second embodiment lend themselves to an efficient implementation on parallel hardware platforms. Specific technical problems experienced in the implementation of the proposed embodiments on commercial radar devices are identified and possible solutions are proposed. Our numerical results lead to the conclusion that the described embodiments are able to generate accurate radar images in the presence of multiple and closely spaced targets.

Low cost commercial radar devices can be used to implement the present invention.

State of the art

MIMO radar systems can be divided in statistical MIMO radars and colocated MIMO radars on the basis of the distance between their transmit array and their receive array; in the first case, transmit and receive antennas are widely separated, whereas, in the second one, they are closely spaced and, in particular, they are usually placed on the same shield. In this document, we focus on colocated MIMO radars only; such systems play an important role in a number of applications, because of their limited cost, their small size and their ability to detect the presence of multiple targets. The performance achieved by any colocated MIMO radar system depends not only on some important characteristics of its hardware (e.g., the operating frequency, the number of transmit and receive antennas, the configuration of the transmit and receive arrays, etc.), but also on the techniques that are employed in the generation of its radiated waveforms and in the processing of the measurements acquired through its receive array. As far as the last issue is concerned, it is worth stressing that optimal (i.e., maximum likelihood, ML) techniques for the estimation of the overall number of targets and of their spatial coordinates cannot be employed in practice, since they require solving complicated multidimensional optimization problems and, consequently, entail a huge computational effort, even in the presence of a small number of targets. This has motivated the development of various sub-optimal estimation techniques able to achieve good estimation accuracy at a manageable computational cost. A well known sub-optimal technique employed in real world radar systems is the one described in "S. Rao, "MIMO Radar," Texas Instruments Application Report SWRA554A, July 2018, for FMCW radar systems".

It requires: a) the computation of a multidimensional Fast Fourier Transform (FFT) of the real matrix collecting the time- domain samples of the signals acquired through the receive array of the employed radar device; b) the search for the peaks of the resulting amplitude spectrum over a range-azimuth-elevation domain or a range-azimuth domain in 3D and 2D imaging, respectively. Despite the practical importance of this technique, the computational effort it requires is still significant, since it involves multidimensional spectral analysis of the acquired signals. Moreover, it suffers from the following relevant drawback: it can miss targets whose electromagnetic echoes are weaker than those generated by other spatially close targets; this is due to the fact the spectral contribution due to weak echoes is usually hidden by that produced by stronger echoes. This drawback may substantially affect the overall quality of radar imaging in the presence of extended targets, since such targets can be usually modelled as a cluster of point targets characterized by different radar cross sections.

Alternative sub-optimal techniques available in the technical literature are

- D. Oh and J. Lee, "Low-Complexity Range-Azimuth FMCW Radar Sensor Using Joint Angle and Delay Estimation Without SVD and EVD," IEEE Sensors Journal, vol. 15, no. 9, pp. 4799-4811, Sep. 2015 and J.-J. Lin, Y.-P. Li, W.-C. Hsu, and T.-S. Lee, "Design of an FMCW radar baseband signal processing system for automotive application," SpringerPlus, vol. 5, no. 42, pp. 1-16, Jan 2016 and

- E. Sirignano, A. Davoli, G. M. Vitetta, and F. Viappiani, "A Comparative Analysis of Deterministic Detection and Estimation Techniques for MIMO SFCW Radars," IEEE Access, vol. 7, pp. 129848-129 861, 2019.

They are based on the idea of turning a complicated multidimensional optimization problem into a series of simpler and interconnected optimization sub-problems, each of which involves a search for the local maxima of a specific cost function over a limited one-dimensional (ID) or 2D parameter space; this idea is exploited to develop range-azimuth estimators for frequency modulated continuous wave (FMCW) MIMO radars and for stepped frequency continuous wave (SFCW) MIMO radars. More specifically, target delays are first estimated by applying the multiple signal classification (MUSIC) algorithm to a temporal auto-correlation matrix or by identifying the beat frequencies in the downconverted received signal through spectral analysis (more specifically, through the FFT algorithm) in the solution proposed by Oh and in that proposed by Lin, respectively; then, the acquired information are exploited to estimate the DOA of the echoes originating from the detected targets. In the solutions developed by Sirignano, instead, the parameters of the dominant target are estimated first; then, its contribution to the received signals is cancelled and the resulting residual signals are processed to detect the next dominant target. This detection/estimation/cancellation procedure is iterated until no significant target is detected on the basis of the residual signals. It is desired to improve the results obtained by Oh, Lin and Sirignano.

It should be clear that the FFT represents the practical implementation of the DFT through computer means. Therefore, DFT and FFT can be used in an interchangeable way. The same concept applies also to IFFT and IDFT.

Summary of the invention

The present solution shares some important features with the detection/estimation techniques developed in Lin and Sirignano. Indeed, similarly as the techniques illustrated in Sirignano, the present method is deterministic, it processes a single snapshot, operates in an iterative fashion and is computationally efficient; its last feature can be related to the fact it requires the evaluation of ID FFTs only and the search for the global maximum of proper cost functions over ID (frequency, azimuth or elevation) domains. Moreover, similarly to Lin, it first extracts the range of each detected target from the spectra of the received signals and, then, fuses the information originating from such spectra to extract direction of arrival (DOA) information. However, unlike the methods developed in prior art, it can be employed in both FMCW and SFCW radar systems.

Moreover, its two embodiments exploit new techniques for estimating the parameters of multiple overlapped sinusoids or multiple overlapped complex exponentials in the presence of additive noise.Advantageously, the present solution involve the following concepts:

1. The technique employed in range estimation computes three distinct ID FFTs of the baseband signal acquired through all the receive antennas or through a portion of them. Moreover, it exploits a serial cancellation algorithm to make the detection of closely spaced targets possible. The use of multiple FFTs allows to achieve high accuracy in range estimation. This, in turn, results in a mitigation of the error accumulation phenomenon affecting, like any iterative deterministic method, the serial cancellation procedure on which the detection of multiple targets is based. Consequently, a good resolution is achieved, even in the presence of a large number of targets.

2.The technique employed in DOA estimation is conceptually similar to the one developed for range estimation, but processes the measurements referring to each of the ranges at which the presence of at least one target has been detected. For this reason, it accomplishes ID FFT processing in the estimation of the angular coordinates of the target/targets detected at a specific range. Moreover, it exploits a serial cancellation algorithm to make the detection of such targets possible, even in the presence of similar angular coordinates, and a new method, called spatial folding, to detect ghost targets and refine range estimates. The combination of these algorithms allows to solve the problem of merged measurements or unresolved measurements, in the sense that targets generating merged measurements in the range domain are resolved at this stage.

The present invention will be made fully clear with the aid of the attached dependent claim describing preferred embodiment of the invention.

Brief description of the figures

The invention will become fully clear through the following detailed description, given by way of a mere exemplifying and non limiting example, to be read with reference to the attached drawing figures, wherein:

Figure 1 shows an example of architecture of a MIMO FMCW radar system;

Figure 2 discloses the time diagram of the instantaneous frequency of the RF signal radiated in a FMCW radar system;

Figure 3 shows an example of architecture of a MIMO SFCW radar system; Figure 4 discloses the time evolution of the instantaneous frequency of the RF signal radiated in a SFCW radar system;

Figure 5 discloses a block diagram of a preferred embodiment of the proposed method;

Figure 6 discloses a virtual antenna array considered in the proposed method;

Figure 7 discloses a block diagram describing the first embodiment of the proposed method; Figure 8a and 8b disclose: a) a representation of the vector |X_{V j}|; and b) a discrete range profile computed by the single target detection, range estimation and cancellation STDREC algorithm, subject of the present invention, respectively;

Figure 9 discloses the identification of the reference antenna on the considered virtual array;

Figure 10 shows an example of a reference vertical uniform virtual antenna (VULA) and a reference horizontal uniform virtual antenna (HULA) including the reference antenna;

Figure 11 shows the representation of the vertically folded HULAs;

Figure 12 shows a block diagram representing a second embodiment of the proposed method;

Figure 13 and Figure 14 disclose a physical array used in our experimental work and the corresponding virtual array associated with the physical array, respectively;

Figure 15 discloses a block diagram representing the first embodiment of the proposed method; a compensation technique based on deep-learning methods is employed in the SPE; Figure 16 discloses a block diagram representing the proposed deep network architecture; Figure 17 discloses a diagram of the Amplitude spectra of the received signal and of the residual and the estimates of the amplitudes computed by the STDREC algorithm; Figure 18 discloses a 3D reconstruction of the propagation Scenario (radar image); Figure 19a and 19b disclose the Amplitude of the elements of the vector and of the vector and the estimates of the amplitudes computed by the STDAEC algorithm; Figure 20 discloses a 3D reconstruction (radar image) of the Scenario #3 for both FMCW and SFCW radar system running embodiment #1; Figure 21 shows a block diagram representing the RPE in a FMCW radar system; Figure 22 shows a block diagram of the SFE#1 algorithm employed in a FMCW radar system for range estimation; Figure 23 shows a block diagram of the SFE#2 algorithm employed in a FMCW radar system for range estimation; Figure 24 shows a block diagram of the CSFE#1 algorithm employed in a FMCW radar system for range estimation;

Figure 25 shows a block diagram of the CSFE#2 algorithm employed in a FMCW radar system for range estimation;

Figure 26 shows a block diagram of the fusion step in RPE;

Figure 27 shows a block diagram representing the SPE in a FMCW/SFCW radar system;

Figure 28 shows a block diagram of the acquisition step in SPE;

Figure 29 shows a block diagram of an iteration of the STDAEC algorithm;

Figure 30 shows a block diagram of the CSFE#1 algorithm employed in a FMCW radar system for vertical frequency estimation; Figure 31 shows a block diagram of the CSFE#2 algorithm employed in a FMCW radar system for vertical frequency estimation;

Figure 32 shows a block diagram of the CSFE#1 algorithm employed in a FMCW radar system for horizontal frequency estimation; Figure 33 shows a block diagram of the CSFE#2 algorithm employed in a FMCW radar system for horizontal frequency estimation;

Figure 34 shows a block diagram of the CSFE#1 algorithm employed in a FMCW radar system for the refinement step;

Figure 35 shows a block diagram of the CSFE#2 algorithm employed in a FMCW radar system for the refinement step;

Figure 36 shows a block diagram representing the RPE in a SFCW radar system;

Figure 37 shows a block diagram of the CSDE#1 algorithm employed in a SFCW radar system for range estimation;

Figure 38 shows a block diagram of the CSDE#2 algorithm employed in a SFCW radar system for range estimation;

Figure 39 shows a block diagram of the CSDE#1 algorithm employed in a SFCW radar system for vertical frequency estimation;

Figure 40 shows a block diagram of the CSDE#2 algorithm employed in a SFCW radar system for vertical frequency estimation;

Figure 41 shows a block diagram of the CSDE#1 algorithm employed in a SFCW radar system for horizontal frequency estimation; Figure 42 shows a block diagram of the CSDE#2 algorithm employed in a SFCW radar system for horizontal frequency estimation; Figure 43 shows a block diagram of the CSDE#1 algorithm employed in a SFCW radar system for the refinement step;

Figure 44 shows a block diagram of the CSDE#2 algorithm employed in a SFCW radar system for the refinement step.

The description of the propagation Scenarios #1, #2 and #3 is compared with that provided by their estimates through the attached tables 1 - 6, at the bottom of the detailed description. Table 1. Target parameters employed in scenario #1 and their estimates.

Table 2. Target parameters employed in scenario #2 and their estimates. Table 3. Target parameters employed in scenario #3 and their estimates obtained with the FMCW MIMO radar and embodiment #1.

Table 4. Target parameters employed in scenario #3 and their estimates obtained with the FMCW MIMO radar and embodiment #2.

Table 5. Target parameters employed in scenario #3 and their estimates obtained with the SFCW MIMO radar and embodiment #1.

Table 6. Target parameters employed in scenario #3 and their estimates obtained with the SFCW MIMO radar and embodiment #2.

Table 7. RMSE and peak errors for scenario #3.

The same reference numerals and letters in the figures designate the same or functionally equivalent parts.

According to the present invention, the term "second element" does not imply the presence of a "first element", first, second, etc., are used only for improving the clarity of the description and they should not be interpreted in a limiting way.

Detailed description of preferred embodiments of the invention 1. Radar system and signal models

In this section, the architecture of MIMO FMCW and SFCW colocated radar systems is illustrated. Moreover, the mathematical models of the signals transmitted and received in both systems are illustrated. 1.1. MIMO FMCW radar system

In this paragraph, we take into consideration the colocated MIMO FMCW radar system illustrated in Fig.l; we assume that this system is equipped with a 2D array. The radio frequency (RF) signal radiated by the radar transmitter is generated as follows. A ramp generator feeds a voltage controlled oscillator (VCO), that produces a chirp frequency modulated signal, whose instantaneous frequency evolves periodically, as illustrated in Figure 2; in this figure, the parameters T, T_R and T₀ represent the chirp interval, the reset time and the pulse period (or pulse repetition interval), respectively , whereas f₀ and B are the so called start frequency and the bandwidth, respectively, of the frequency modulated signal. The FM signal feeds a power amplifier whose response is sent to one of the N_T available transmit (TX) antennas. It can be assumed that a time division multiplexing (TDM) strategy is employed in the selection of the TX antennas. This means that all the available TX antennas are used over N_T consecutive chirp intervals, i.e. in a single transmission frame (lasting T_F = N_TT₀ s), so that all the available transmit diversity is exploited. It is worth stressing that, as already mentioned above, the proposed method processes a single snapshot of the received signals, i.e. the signals received in a transmission frame. Moreover, it relies on the assumption that the propagation scenario in which the radar system operates undergoes negligible variations in the observation interval, i.e. over ^_^ s. For this reason, the presence of the Doppler phenomenon is not accounted for in the received signal model illustrated below. The RF signal radiated in a single chirp interval and, in particular, in the time interval (0, ^), can be expressed as (see Figure 1) where is the signal phase, ^_^^ is its amplitude is the chirp rate, i.e. the steepness of the employed frequency chirp. In the following we assume that the radiated signal ^_^^(^) (1.1) is reflected by ^ targets and that the position of the ^- th target (with ^ = 0, 1, ..., ^ − 1) with respect to the center of the antenna array is identified by its range ^_^, its azimuth ^_^ and its elevation ^_^. Consequently, the useful component of the signal captured by each of the ^_^ receive (RX) antennas consists of the superposition of ^ echoes, each originating from a distinct target. Moreover, in deriving our method for detecting such targets and estimating their spatial coordinates, we usually refer to the ^-dimensional column vector of the baseband signal samples acquired through the RX virtual antenna (RVA) identified by the horizontal index ^ and the vertical index ^. In the remaining part of this document, this antenna is denoted (^, ^) RVA or ^-th RVA (with ^ = 0, 1, ..., ^_^^ − 1), for brevity, where ^ is a single index related to the couple (^, ^) by a specific one-to-one mapping. If the single index ^ is adopted, the notation ) is employed for the vector . It is important to point out that the ^-th RVA virtual antenna is associated with a specific couple (TX antenna, RX antenna) of the 2D physical antenna array and represents a single element of the 2D virtual antenna array, which, generally speaking, consists of elements. Moreover, if the considered radar is equipped with a uniform rectangular array (URA), it can be assumed that its virtual array is made of horizontal uniform linear arrays (ULAs), each consisting of elements, so that

N_el. Additional details about this technical issue can be found, for example, in Sirignano.

The expression of the n-th element of the received signal vector can be derived as follows.

It can be shown that the delay of the echo generated by the Z-th target and observed by the v-th RVA is (e.g., see eq 4 in "I. Bilik, 0. Longman, S. Villeval, and J. Tabrikian, "The rise of radar for autonomous vehicles: Signal processing solutions and future research directions," IEEE Sig. Proc. Mag., vol. 36, no. 5, pp. 20-31, Sep. 2019") where the elements of the couple (x_v, y_v) are the coordinates of the considered RVA with respect to a Cartesian reference system lying on the physical antenna array (the center of the virtual array is identified by (x_v,y_v) = (0,0)) . The RF signal originating from this target is amplified by a low noise amplifier (LNA) and, then, undergoes a frequency downconversion based on the locally generated chirp FM signal and band-pass filtering; if the delay due to this processing is neglected¹, the phase of the

¹ Note that this assumption makes sense if all the RX signal chains introduce the same delay. Unluckily, this does not hold in commercial radar devices. However, the delays of all the signal chains can beaccurately estimated and their differences, being deterministic quantities, can be accurately compensated for by means of an accurate calibration of the employed radar device. resulting beat signal (i.e., of the downconverted signal) available at the output of the band-pass filter is where and since the term is usually negligible with respect to the term Therefore, the overall beat signal originating from the can be expressed as for where and t_M denote the minimum and the maximum of the target delays is the amplitude of the Z-th component of the useful signal (this amplitude depends on both the range and the reflectivity of the Z-th target) and is zero mean additive noise affecting the considered RX signal. The signal undergoes analog-to-digital conversion at a frequency where denotes the sampling period. If, for simplicity, quantization effects are neglected, sampling the signal at the instant results in the real quantity that represents the n-th element of the vector (with n = for any n and N denotes the overall number of samples acquired from each RVA and

(1.13)

In our computer simulations, for any v, the sequence { w_{v n}} consists of zero mean and independent Gaussian random variables, each having variance

In some FMCW radar devices, the frequency downconversion of the received RF signals entails the extraction of both the inphase and quadrature components of each of such signals. When this occurs, eqs. (1.11) and (1.12) are replaced by and respectively. The meaning and expression of the parameters appearing in the last two equations are the same as those illustrated for eqs. (1.11) and (1.12), the only difference being represented all the involved signals and their samples are complex (in particular, the sequence consists of zero mean and independent complex Gaussian random variables, each having variance

It is also worth mentioning that the maximum target delay (i.e., the maximum resolvable frequency) depends on the sampling frequency f_s adopted in the analog-to-digital conversion since aliasing must be avoided; more specifically, it is not difficult to show that the inequality

(1.16) must be satisfied. The signal model (1.12) can be also rewritten as where, for any l and v,

(1.18) is the complex amplitude of the Z-th target detected on the RVA and is the normalised frequency associated with the frequency f_Vyi .

It is also important to point out that all the vectors {iy} are generated on the basis of the samples resulting from the analog- to-digital conversion of the N_R downconverted signals associated with all the available physical RX antennas. The samples acquired from each RX antenna within the same chirp undergo serial-to- parallel conversion in the S/P block shown in Figure 1. We assume that this block also produces all the vectors { r_v}, which become available for further processing. 1.2. MIMO SFCW radar system

In this paragraph, we take into consideration the colocated MIMO SFCW radar system illustrated in Figure 3; we assume that this system, like the one shown in Figure 1, is equipped with a 2D array. The signal it radiates is generated as follows. A staircase waveform generator feeds a VCO, that produces a modulated continuous wave; this signal feeds a power amplifier whose response is sent to one of the N_T available TX antennas. In the following, the parameters T, T_R and T₀ represent the duration of each staircase, the reset time and the repetition interval, respectively, as illustrated in Figure 4. Moreover, similarly as the previous radar system, a TDM strategy can be employed in the selection of the TX antennas. This means that all the available TX antennas are used over N_T consecutive staircase intervals, i.e. in a single transmission frame (lasting T_F = N_TT₀ s), so that all the available transmit diversity is exploited. It is worth stressing that, as already mentioned above, the proposed method process a single snapshot of the received signals, i.e. the signals received in a single transmission frame. Moreover, it relies on the assumption that the propagation scenario in which our radar system operates undergoes negligible variations in the observation interval, i.e. over T_F s. For this reason, the presence of the Doppler phenomenon is not accounted for in the received signal model illustrated below.

The radiated RF signal s_RF(t) is still expressed by eqs. (1.1)—

(1.2). However, eq. (1.3) does not hold any more, since the time evolution of the instantaneous frequency of the signal s_RF(t) is represented by a staircase lasting T s and whose N steps have equal duration; more specifically, the n-th value of this frequency is with where is the minimum radiated frequency, is the frequency step size and N is the overall number of transmitted frequencies (see Figure 4).

The transmitted signal allows to sound the communication channel between the TX and each of the RX antennas at the N equally spaced frequencies expressed by eq. (1.20). For this reason, the processing accomplished at the receive side of the radar system aims at generating an estimate of the channel frequency response at each of these frequencies. More specifically, these estimates are collected in the vector

(1.21) that refers to the RVA identified by the horizontal index p and the vertical index q (and denoted RVA or 17-th RVA); the equivalent notation can be also adopted, where v is a single index related with the couple (p,q) through a specific one-to-one mapping. It is not difficult to prove that, under the same assumptions made in the derivation of eq. (1.12), the n-th element x_vn of the last vector can be expressed as

(1.23) with where

(1.24) and the parameters and and the random variable have exactly the same mea ning as the one illustrated for the received signal model (1.12)(in particular, t_nί is still expressed by eq. (1.7)). In our computer simulations, for any v, the sequence {^wv,n} consists of zero mean and independent complex Gaussian random variables, each having variance Moreover, the first term appearing in the right-hand side (RHS) of eq. (1.23) represents the frequency response of the communication channel associated with the v-th RVA at the frequency f_n (1.20). It is also worth noting that eq. (1.23) can be put in the form where i A a_lexp(_^—jp_{v l}) and

(1.27)

T_v,i = ^tn,i D/ represent the complex amplitude characterizing the contribution of the Z-th target to x_{v n} (1.25) and the delay of that target normalised with respect to 1/D/, respectively. The model (1.25), similarly as the model (1.12) developed for a EMCW radar, allows to interpret the signal observed on the v-th RVA as a superposition of L oscillations. However, the former model is complex, whereas the latter one is real. This difference has to be kept carefully into account in our description of the detection/estimation techniques illustrated in the following section.

2. Description of the Proposed Method Proposed and of Its Two Embodiments In this section, a novel method for detecting multiple targets and estimating their spatial coordinates (i.e., for generating a 3D or 2D radar image) in colocated MIMO FMCW/SFCW radar systems is described. First, we provide a general description of this method. Then, we illustrate two specific embodiments of it for a FMCW radar system employed for 3D or 2D radar imaging. Finally, we discuss how they can be adapted to a SFCW radar system.

2.1. General description of the proposed method

The overall architecture of the method we propose for the detection of multiple targets and the estimation of their spatial coordinates in a colocated MIMO radar system is described by the block diagram illustrated in Figure 5. Our method operates as follows. First of all, each vector of the set {iy} ({x_v}; see eqs. (1.6) and (1.22), respectively) undergoes FFT processing, so that signal analysis is moved from the time-domain (frequency- domain) to the frequency-domain (time-domain) in a FMCW (SFCW) radar system. The FFT output is processed by a range profile estimator (RPE); this generates an estimate of the target range profile (TRP), i.e. it identifies each range at which relevant echoes are detected and estimates their energy; the last quantity allows us to rank each identified range on the basis of its perceptual importance. The output of the FFT processing and the target range profile are processed by a spatial estimator (SPE). This block estimates the number of targets contributing to the range profile and their angular parameters (i.e., azimuth and elevation); moreover, it eventually produces a finer estimate of their range. Its output is represented by the set ; 1= 0, 1, ..., L— 1}, where L represents an estimate of the parameter L (i.e., of the overall number of targets), whereas R_ir and f_i represent an estimate of the range R_ir azimuth and elevation fi, respectively, of the Z-th target (with 1= 0, 1, ..., L— 1).

The approach adopted in our method allows to separate range estimation from angular estimation, i.e. to turn a 3D detection/estimation problem in - a ID detection/estimation problem involving the detection of multiple targets and the estimation of their range only, plus

- a 2D estimation problem concerning the detection of targets having the same range and the estimation of their angular coordinates only.

Consequently, the overall problem of detecting multiple targets and estimating their range and angles is turned into a couple of simpler detection/estimation problems. Moreover, in our method, the techniques listed below are exploited to develop the computationally efficient signal processing techniques employed in the RPE and in the SPE.

Target range profile estimator a) Antenna selection - The RPE processes the FFT outputs referring to all the RVAs or to a portion of them (say, N_A of the N_VR RVAS). On the one hand, a larger N_A allows to generate a more accurate estimate of the TRP (i.e., of the number of detected targets, of their ranges and of their energies). On the other hand, the selection of a smaller N_A results in a reduction of the overall computational effort devoted to the estimation of the TRP. b) Parallel processing and data fusion - The measurements acquired through the N_A RVAs mentioned in the previous point are processed in two steps. In the first step, target range estimation is accomplished on each RVA independently of all the other RVAs, i.e. the acquired measurements are processed on a per-antenna basis; this is beneficial when parallel computing hardware is employed in the execution of the first step. In the second step, the target range information extracted from each of the selected N_A RVAs are fused to generate the TRP. c) Serial target cancellation - Target detection and range estimation on each RVA is a multidimensional problem since it aims at detecting multiple targets and estimating their ranges. In our method, this multidimensional problem is turned into a sequence of ID estimation problems by adopting a serial interference cancellation (SIC) approach (e.g., see "E. Sirignano, A. Davoli, G. M. Vitetta, and F. Viappiani, "A Comparative Analysis of Deterministic Detection and Estimation Techniques for MIMO SFCW Radars," IEEE Access, vol. 7, pp. 129 848-129 861, 2019."). This means that the noisy signal observed on each RVA is processed to detect a single (and, in particular, the strongest) target, and to estimate its range and complex amplitude. Then, the contribution of this target to the noisy signal is estimated and subtracted from the signal itself (i.e., the detected target is treated as a form of interference), so that target cancellation is accomplished. This detection/estimation/cancellation procedure is iteratively applied to the residual noisy signal until its overall energy drops below a given threshold.

Spatial estimator a) Alternating maximization - The estimation of the angular coordinates (namely, azimuth and elevation) of a target whose range is known requires searching for the maximum of a proper cost function over a 2D domain. In our method, the alternating maximization (AM) technique is exploited to turn a multidimensional optimization problem into a set of interacting optimization problems, each involving a single variable and that are solved in an iterative fashion (e.g., see Par. IV-A "I. Ziskind and M. Wax, "Maximum likelihood localization of multiple sources by alternating projection," IEEE Trans Ac., Speech and Sig. Proc., vol. 36, no. 10, pp. 1553-1560, 1988."). More specifically, this technique is employed to develop iterative algorithms alternating the estimation of the elevation of a given target with that of its azimuth; for this reason, a 2D optimization problem is turned into a couple of interacting ID optimization problems. b) Serial target cancellation - Each of the ranges collected in the TRP is associated with an unknown number of targets; for this reason, the SPE is required to resolve all the targets associated with a given range and estimate their angular coordinates. In the technical literature about radar systems, the detection of an unknown number of targets characterized by the same range (or by ranges whose difference is below the range resolution of the employed radar system) and the estimation of their angular parameters is known to be a difficult multidimensional problem (e.g., see Par. III-C "D. Oh and J. Lee, "Low-Complexity Range-Azimuth FMCW Radar Sensor Using Joint Angle and Delay Estimation Without SVD and EVD, " IEEE Sensors Journal, vol. 15, no. 9, pp. 4799-4811, Sep. 2015."). In our method, this multidimensional problem is turned into a sequence of 2D estimation problems by adopting a SIC approach (see "E. Sirignano, A. Davoli, G. M. Vitetta, and F. Viappiani, "A Comparative Analysis of Deterministic Detection and Estimation Techniques for MIMO SFCW Radars," IEEE Access, vol. 7, pp. 129 848-129 861, 2019." and references therein). This means that the noisy signal referring to a specific range and acquired on all the RVAs is processed to detect a single (and, in particular, the strongest) target, and to estimate its angular coordinates and its complex amplitude. Then, the contribution of this target to the noisy signal is subtracted from the signal itself (i.e., the detected target is treated as a form of interference), so that target cancellation is accomplished. This detection/estimation/cancellation procedure is iteratively repeated on the residual noisy signal until its overall energy drops below a given threshold. Moreover, in our method, the SIC technique is combined with the AM approach described in the previous point; this allows to detect and estimate the angular parameters of a single target (i.e., to solve a 2D optimization problem) by means of an iterative procedure alternating the estimation of its elevation with that of its azimuth (i.e., by means of a technique solving a couple of ID optimization problems). c) Parallel processing and data fusion - Two different approaches can be used in the detection and estimation of multiple targets characterized by different ranges. In fact, the detection and the estimation of the targets associated with distinct ranges appearing in the TRP can be accomplished in a parallel fashion or in a sequential fashion. The first approach is potentially very useful when spatial estimation is executed on parallel computing hardware. In fact, in this case, multiple spatial estimation algorithms can be run in parallel, one for each of the ranges appearing in the TRP. Note, however, that target information generated by these algorithms need to be fused at the end of their processing. In fact, as it will become clearer in the following, spatial estimators analysing the measurements that refer to close ranges contained in the TRP may lead to the detection of the same target; in other words, some targets can be detected multiple times. The second approach is based on the idea of sequentially analysing the measurements associated with each of the ranges appearing in the TRP. In doing so, the order to be followed is based on the perceptual importance of these ranges, i.e. on their energy. Two specific embodiments of the general method illustrated above are described in the following paragraph. In both cases, we assume, without loosing generality, that the available measurements are acquired through the N_VH X N_vv virtual URA illustrated in Figure 6; the horizontal (vertical) spacing between adjacent antennas in this array is denoted d_{VH (}d_vv) . Moreover, the embodiments are described first for a FMCW radar system employed for 3D imaging; then, we briefly discuss how they can be modified to be employed in a SFCW radar system and in the case of 2D radar imaging (see Paragraphs 2.3. SFCW radar system and 2.4. Adaptation to 2D Imaging, respectively).

2.2. FMCW radar system

In summary, in the present invention is presented a method for the detection of targets through a MIMO FMCW radar equipped with a plurality of transmitting (TX) and receiving (RX) antennas, wherein each couple of said TX and RX antennas is replaced with the equivalent virtual antenna; said MIMO FMCW radar is arranged to generate real or complex signals in response to a propagation scenario including a plurality of point targets; said method includes (see Steps 1 - 6 of Figure 21):

— (v)

- (Step 1: FFT processing) estimation of the spectrum (X₀ ) of

— (v) said signal and its first N_FFT — 1 derivatives ({X_fc ;k = 1,...,N_FFT — 1}) with N_FFT = 3 when said signal is real and N_FFT = 4 when said signal is complex, for each of said virtual antennas through FFT calculation with an oversampling factor (M),

- (Step 2: RVA selection) acquisition of a sub-set of a number (N_a) of virtual antennas spectrum and of the corresponding derivatives,

Recursive execution of (Step 3: STDREC-S1) calculation of an estimate of the parameters of the most dominant point target, including phase (p), amplitude (a) and frequency (/) for each of said virtual antennas of said sub-set on the basis of said spectrum, and its derivatives, through N_FFT — 1 FFTs for each of said virtual antennas of said sub-set,

- (Step 4: STDREC-S2) cancellation of said most dominant point target and computation of a residual spectrum,

- (Step 5: STDREC-S3) computation of an energy of said residual spectrum and return to (Step 3) until said energy is over a predetermined threshold (TSTDREC)_r otherwise

- (Step 6) computation of an average energy and fusion of ranges information of said point targets detected by said number of virtual antennas to list said point targets according to said average energy in a decreasing order.

2.2.1. Embodiment #1

The processing accomplished in the first embodiment of the proposed method is described by the block diagram illustrated in Fig. 5 and can be divided in three tasks, each associated with one of the blocks appearing in that figure (the i-th task is denoted Ti in the following).

In all these tasks, the signal received over a single array snapshot is processed. In the following, the structure of each task is described; additional details about the employed algorithms are provided in Paragraph 4.

Tl) FFT processing - The processing accomplished within this task can be summarized as follows. Given the vector r_v (1.6), the iV-dimensional vectors (2.1) and are generated for v = 0, 1, ..., N_{VR —} 1; here,

(2.3) with h = 0, 1, ..., N — 1 and m = 1, 2. Then, the vectors r_v, x_lv and Cc_n undergo zero padding (ZP) for any v; this produces the iV₀- dimensional vectors

(2.4) and

(2.6) where M is a positive integer (dubbed oversampling factor), 0_D is a D-dimensional (column) null vector and

(2.7)

N₀ = M ^■ N. Finally, the N₀-dimensional vectors with m = 0,1,2 are computed for any v (i.e., for any p and q); here, DFT_WQ[X] denotes, up to a scale factor, the N₀-th order discrete Fourier transform (DFT) of the iV₀-dimensional vector X. More specifically, in the following, we assume that

(2.9) with k = 0, 1, ..., iV₀ — 1 and m = 0,1,2. The set {X_0iV, Xi_,v, X_2,v' ^{v =}

0, 1, ..., N_VR — 1}, consisting of 3·N_VR N₀-dimensional vectors, is passed to both the RPE and SPE blocks, whose objectives have been summarized in the previous section. It is important to point out that: a) For any v, each of the vectors {X_0jV, X_lv, X-2_,v] (see eq. (2.8)) can be computed by executing a N₀-th order FFT. b) The vector X_0/y (2.8) consists of N₀ equally spaced samples of the spectrum of the sequence {r_vn} acquired on the v-th RVA (see eq. (1.12)). The vectors X_ljV and X_2v (see eq. (2.8)) instead, consist of, up to a scale factor, N₀ equally spaced samples of the first derivative and of the second derivative, respectively, of the same spectrum. c) As already illustrated in the previous paragraph, the RPE aims at computing a preliminary estimate of the ranges (i.e., of the frequencies) at which the energy reflected by most of the targets is concentrated, i.e. at accomplishing range profile estimation. On the other hand, the SPE aims at estimating the complete spatial coordinates of the targets associated with each of the ranges included in the TRP, i.e. at accomplishing spatial estimation. The SPE exploits the range information generated by the RPE in order to concentrate its computational effort on a set of well defined ranges; this allows to reduce the dimensionality of the search space involved in spatial estimation. This explains also why the processing accomplished by the SPE cannot start before that at a least a portion of the range/energy information generated by the RPE becomes available.

T2) RPE - The aim of this task is generating a preliminary estimate of the number of targets and of their ranges. Note that, in this step, multiple targets having the same range or ranges whose difference is below the range resolution of the employed radar system are detected as a single target and no effort is made to separate their contributions. The processing accomplished within this task involves N_A of the N_VR RVAs and consists of three consecutive steps listed below (the i-th step is denoted T2-Si in the following); each step is associated with one of the blocks contained in the RPE block appearing in Figure 7.

T2-S1) RVA selection - In this step, N_A distinct RVAs are selected within the set of the N_VR available RVAs on the basis of a specific algorithm. Consequently, only a portion of the set of triads {(Xo_{,V /}· Xi _iV , X-2_,v)t v = 0, 1, ..., N_VR — 1} is processed in this step. In the following, the selected triads are denoted {(Xo.„, ^xi_,vr ^x _2,i0; VES_A}, where S_A = {v₀, v_lr ..., i¾ _i} is the set of the values of the index v identifying them. T2-S2) Target detection and range estimation - This step aims at detecting the most relevant targets on each of the N_A antennas selected in the previous step and at estimating their ranges (i.e., the frequencies associated with these ranges; see eqs. (1.7) and (1.9)) and their complex amplitudes (see eq. (1.18)).

The processing accomplished within it is executed on an antenna- by-antenna basis and exploits a novel iterative estimation algorithm called single target detection, range estimation and cancellation (STDREC) and whose detailed description is provided in the Paragraph 4.1. The STDREC processing for the v_k-th RVA (with k = 0, 1, N_A — 1) can be summarized as follows. First of all, this algorithm is initialised by setting

(2.10) g(°) g

^L m,v_k — ^L m,v_k> with m = 0,l,2 and the iteration index i to 1. Then, it starts executing its iterations; moreover, in its i-th iteration, it accomplishes the three steps described below to detect a new target and cancel its contribution to the triad

(the p-th step of the considered i-th iteration is denoted STDREC-Sp in the following).

STDREC-S1) Detection of a new target and estimation of its parameters - The triad -'-^s processed to detect a new (i.e., the i-th) target, and to estimate the frequency and the complex amplitude C® associated with it. Note that, generally speaking, the frequency is not a multiple of the fundamental frequency (2.11) associated with the FFT processing executed in task Tl ; for this reason, generally speaking, it can be expressed as

(2.12) where a® is an integer parameter belonging to the set {0, 1, N₀ — 1} and is a real parameter, such that

(0£d® £0.5 if <^} = 0 j-O.SCS^ O ifa® =N₀-1.

1 —0.5 < d,^vRk < 0.5 otherwise The processing accomplished in this step is executed by an algorithm, dubbed single frequency estimator (SFE) if the received sequence {r_vn} is real and complex single frequency estimator (CSFE) if the received sequence {r_vn} is complex; this algorithm represents the most important part of the STDREC algorithm and is described in detail in Paragraph 4.1.2 if the received sequence {r_vn} is real and in Paragraph 4.1.3 if the received sequence {r_vn} is complex. The SFE/CSFE computes the estimates of the parameters C®, and f_v^' respectively, on the basis of the triad (X¾_fc, X¾_fc, X¾_fc).

In summary, the above method can be described as follows: when a base-band signal made available by said radar is real in phase-component only, Step 3 of Figure 21 includes (see Steps 3.1.0 - 3.1.2 of Figure 22):

- (Step 3.1.0) representation of a target normalized frequency F = f/f_s, where f_s is the sampling frequency, as sum of a main portion (aF_dft) and a residual portion (5F_dft), where F_DFT = l/iV₀, where N₀ is the FFT order, wherein said main portion is

_L — (v) determined by searching the index (a) of the vector element X_0,a corresponding to the maximum absolute value in a first half plus

— (v) one elements of the same vector X₀ and said residual frequency is found through an iterative procedure together with the corresponding target complex amplitude C= aexp(jxp)/2, where a is said target amplitude and y is said target phase.

- (Step 3.1.1) Said iterative procedure includes a preliminary initialization comprising: a) setting the iteration index i to 1, the initial estimate of

— (v) target complex amplitude C equal to said vector element (X_oa)_r the initial estimate of D = 2pdF_DFT to zero and the initial estimate of F to said main portion (CCF_dft); b) computing the complex coefficients (K_p(2a);p = 1,2,3} as the 2a- th element of the iV₀-order DFT of the sequence n^p, with p = 1,2,3; then iteratively - (Step 3.1.2) the i-th iteration is fed by the estimates of D, F and C computed in the previous iteration, to generate new estimates of D, F and C, for a predetermined number of iterations, wherein each iteration includes a) a first step where the residual D is calculated solving the first order equation Z¾D + c¾ = 0 OR the second order equation a_dD² + ί¾D + C_Q ⁼ 0 in the variable D with: where ¾{x} denotes the real part and 3{x} denotes the imaginary part of the complex quantity x; discarding one of the two solutions, which does not satisfy the inequalities: ifa = 0 otherwise if said second order equation is solved; then the new estimate of F is obtained by adding said residual (AF_dft) to said main portion (af_DFT); b) a second step wherein said new estimate of F is employed to compute the new estimate of C as is generated by

— (v) interpolating the elements of said vector MX₀ where M is said oversampling factor employed in said FFT calculations.

The above mentioned Steps 3.1.1 and 3.1.2 of Figure 22 can be replaced by an alternative method including (see Steps 3.2.1 - 3.2.2 of Figure 23):

- (Step 3.2.1) Said iterative procedure includes a preliminary initialization comprising: a) setting the iteration index i to 1, the initial estimate of C (v) equal to said vector element (Xo_,a)_r the initial estimate of A = 2ir<5F_DFT to zero and the initial estimate of F to said main portion (^QCFDFT)r b) computing the complex coefficients (K_p(2a);p = 1,2,3} as the 2a- th element of the iV₀-order DFT of the sequence n^p, with p = 1,2,3; c) computing the real quantity p=F/F_DFT; then iteratively

- (Step 3.2.2) the i-th iteration is fed by the estimates of D, F and C and the real quantity p computed in the previous iteration, to generate new estimates of D, F, C and p, for a predetermined number of iterations, wherein each iteration includes a) a first step where the residual D is calculated solving the first order equation Z¾D + c¾ = 0 OR the second order equation in the variable D with: where ¾{x} denotes the real part, 3{x} denotes the imaginary part of the complex quantity x, {K_r(2b); p = 1,2,3} are computed as:

— (^"nΊ with p = 1, 2, 3 and {X_k,p> k = 1,2} as:

(v) where x_{r n} is the n-th element of the said real base-band signal

— (v) acquired on the 17-th antenna; said coefficients ({Xk.p-'k = 1,2}) can be also computed by interpolating the elements of said spectrum

— (v)

(X_fc ); discarding one of the two solutions, which does not satisfy the inequalities: if a = 0 otherwise if said second order equation is solved; then the new estimate of F is obtained by adding said residual (D ¾_RT) to said estimate of F; b) a second step wherein the new estimate of p is computed by dividing said new estimate of F by FDFT/ c) a third step wherein said new estimate of F is employed to compute the new estimate of C as: is generated by

— (v) interpolating the elements of said vector MX₀ where M is said oversampling factor employed in said FFT calculations.

When a base-band signal made available by said radar is complex with in phase and quadrature components, Step 3 of Figure 21 includes (see Steps 3.3.0 - 3.3.2 of Figure 24):

- (Step 3.3.0) representation of a target normalized frequency F = f/ f_s, where f_s is the sampling frequency, as sum of a main portion (aF_dft) and a residual portion (5F_dft), where F_DFT = l/iV₀, where iV₀ is the FFT order, wherein said main portion is

_L — (v) determined by searching the index (a) of the vector element X_oa corresponding to the maximum absolute value of the same vector — (v)

X₀ and said residual frequency is found through an iterative procedure together with the corresponding target complex amplitude A = a exp(jxp), where a is said target amplitude and y is said target phase.

- (Step 3.3.1) Said iterative procedure includes a preliminary initialization comprising: setting the iteration index i to 1, the initial estimate of A

— (v) equal to said vector element (X_0,a) _r the initial estimate of D = 2i F_DFT to zero and the initial estimate of F to said main portion (aF_DFT) then iteratively

- (Step 3.3.2) the i-th iteration is fed by the estimates of D, F and A computed in the previous iteration, to generate new estimates of D, F and A, for a predetermined number of iterations, wherein each iteration includes a) a first step where the residual D is calculated solving the first order equation Z¾D + c¾ = 0 OR the second order equation where ¾{x} denotes the real part and 3{x} denotes the imaginary part of the complex quantity x; discarding one of the two solutions, which does not satisfy the inequalities: if a = 0 otherwise if said second order equation is solved; then the new estimate of F is obtained by adding said residual (A F_dft) to said main portion (CC F_dft); b) a second step wherein said new estimate of F is employed to compute the new estimate of A as A = Xi_nt{F) and Xi_nt generated by interpolating the elements of said vector MX₀ where M is said oversampling factor employed in said FFT calculation.

The above mentioned Steps 3.3.1 and 3.3.2 of Figure 24 can be replaced by an alternative method including (see Steps 3.4.1 - 3.4.2 of Figure 25):

- (Step 3.4.1) Said iterative procedure includes a preliminary initialization comprising: a) setting the iteration index i to 1, the initial estimate of A (v) equal to said vector element (X_0@), the initial estimate of A = 2ir<5F_DFT to zero and the initial estimate of F to said main portion

(<2F_DFT)_' b) computing the real quantity p=F/F_DFT; then iteratively

- (Step 3.4.2) the i-th iteration is fed by the estimates of D, F and A and the real quantity p computed in the previous iteration, to generate new estimates of D, F, A and p, for a predetermined number of iterations, wherein each iteration includes: a) a first step where the residual D is calculated solving the first order equation ¾D+ c¾ = 0 OR the second order equation

¾=«(¾ )- where ¾{x} denotes the real part, 3{x} denotes the imaginary part of the complex quantity x and /c= 1,2,3} is computed as: where ay_n is the n-th element of the said complex base-band

— (v) signal acquired on the v-th antenna; said coefficients ({¾ _,p,'k = 1,2,3}) can be also computed by interpolating the elements of said

— (v) spectrum (X_fc ); discarding one of the two solutions, which does not satisfy the inequalities: ifa = 0 otherwise if said second order equation is solved; then the new estimate of F is obtained by adding said residual (AF_dft) to said estimate of F; b) a second step wherein the new estimate of p is computed by dividing said new estimate of F by F_dft; c) a third step wherein said new estimate of F is employed to compute the new estimate of A as s generated

— O) by interpolating the elements of said vector MX₀ where M is said oversampling factor employed in said FFT calculations. STDREC-S2) Cancellation of the new target - The contribution ^Xov_k ^r ^X_iv_k ^r given by the i-th (i.e., by the last) target detected on the i¾-th RVA, to the triad is computed (see eqs. (4.200)-(4.202) if the received sequence { r_vn} is real and eqs. (4.224)-(4.226) if the received sequence { r_vn} is complex) and cancelled from the triad itself. Cancellation consists in the computation of the new residual triad

(2.15) with m = 0,1,2.

In summary, the above method can be summarized by considering that said cancellation, see Step 4 of Figure 21, includes the computation and subtraction of the contribution given by said

—(v) dominant target to said spectrum (X₀ ) and to the corresponding derivatives to compute a residual of the spectrum and of the corresponding derivatives. STDREC-S3) Computation of the residual energy in the frequency domain - The energy of the residual spectrum vector Coϊ^ (2.15) is computed and compared with the positive threshold TSTDREC (which may exhibit a dependence on range). If this energy is below the threshold, the STDREC algorithm stops and L_k = i relevant targets are detected on the v_k -th RVA; otherwise, the recursion index i is increased by one and a new recursion is started (i.e., we go back to

STDREC-S1 ) .

It is worth pointing out that: a) The availability of accurate estimates of the frequency and of the complex amplitude C® plays an important role in this step, since these parameters are exploited in the serial cancellation procedure based on eq. (2.15). In particular, ignoring the fractional part 5® of the frequency (2.12) in this procedure (i.e., assuming that see eq. (2.14)) may result in significant error accumulation. b) The estimate d^ of the fractional part of the frequency (2.12) is exploited in target cancellation, but is not passed to the next (i.e., to the third) step of this task. c) A threshold on the maximum computational effort required by the STDREC algorithm can be set by requiring that the recursion index i never exceeds a fixed threshold; this is equivalent to limit the overall number of targets that can be detected on each RVA. d) The STDREC algorithm generates N_A different data sets; the k- th data set consists of the quadruplets {(a®, 0, 1, ..., L_k — 1}, characterizing the L_k targets detected on the v_k-th antenna (with k = 0, 1, ..., N_A — 1). Note that the overall number of targets may change from antenna to antenna, especially in the presence of extended targets; this is due to the fact that the signals acquired on different RVAs can exhibit significant differences in their structure. e) The following important interpretation of the processing accomplished by the STDREC algorithm on the v_k-th RVA can be given. The vector X_0Vfe ^can be seen as a collection of noisy spectral information referring to iV₀ distinct frequency bins (i.e., to iV₀ distinct range bins) and is usually dense in the presence of multiple extended targets, as illustrated in Figure 8-a) (where the absolute value of its elements is represented). The STDREC allows to extract a discrete frequency (i.e., range) profile from the vector X_0yVk, as illustrated in Figure 8-b). In various real world scenarios, this profile turns out to be sparse, even in the presence of a dense vector X_0Vfe; this is beneficial, since allows to concentrate the RPE computational effort on a set of specific ranges. In the following, the range profile referring to the v_k-th RVA is described by the set of L_k couples i = 0, 1, ..., L_k - 1} (with k = 0, 1, ... ,

N_A — 1); the integer parameter identifies the frequency bin associated with the i-th target (or targets) detected on this

~(£)

RVA, whereas the absolute value of C_Vk represents an estimate of the strength of the echo associated with it. f) The STDREC algorithm can be used for detecting multiple targets and accurately estimating their range in a monostatic radar, i.e., equipped with a single TX antenna and a single RX antenna. g) The STDREC algorithm can be easily extended in a way that multiple targets are detected and estimated in parallel in each of its iterations. If we focus on its i-iteration and the v_k-th RVA, this result is achieved by running multiple (say, instances of the SFE algorithm in parallel. Each of these instances is initialised with the frequency associated with the absolute maximum or a relative maximum detected in the sequence of the absolute values of the elements of the vector . A constraint is set on the minimum spacing between the detected frequencies in order to minimize the interference between the instances running in parallel. Moreover, after identifying the absolute maximum in the above mentioned sequence, a threshold, proportional to such a maximum, is set on the minimum value of the acceptable relative maximum/maxima, so that unrelevant frequencies are discarded. It is also worth stressing that, if a cluster of distinct frequencies is estimated, each of the components of the triplet appearing in eq.

(2.15) consists of the sum of terms, each associated with one of these frequencies.

T2-S3) Range information fusion - This step aims at merging the information provided by the N_A data sets {«S^} generated in the previous step. Its output is represented by the

(2.17)

_<A_b = {a_t;l— 0,1, ...,L_{b —} 1} collecting L_b relevant frequency bins {a_t; 1= 0, 1, ..., L_b — 1} and the set

(2.18)

£_{b —} {E_b,i;I— 0,1, ...,L_{b —} 1} collecting the associated energies; these energies represent the perceptual importance of the identified bins, in the sense that the larger is the energy, the most important is the associated frequency bin. More specifically, the set _<A_b is made of all the integers that appear at least once in the N_A sets

(2.19) where the set

(2.20) 0,1. Lk -i) identifies all the bins detected on the i¾-th RVA. Moreover, the average energy E_bi associated with the ct_j-th bin (with 1= 0, 1, L_b — 1 ) is computed by summing up the square of the absolute value of all the amplitudes such that a® = for any fe 6 {0,

1, N_A — 1} and, for a given k, for any i 6 {0, 1 , ..., L_k — 1}, and dividing it by the overall number of antennas that contribute to this energy (this number is denoted N_bi ); in other words, where and 5[z] = 1 if z = 0 and d[z]= 0 if z ¹ 0 . The final output of the information fusion is represented by the set (2.23) 0,1, ..., L_b — lj.

Note that the cardinality L_b of the last set represents a preliminary estimate of the overall number of targets; this is due to the fact that the energy associated with a given bin may originate from multiple targets, characterized by similar ranges.

In summary, the above fusion step (see Step 6 of Figure21) can be described as follows (see Steps 6.1 - 6.2 of Figure 26):

- (Step 6.1) fusion of the information provided by a number (N_A) of sub-sets collecting all the discretized ranges, named frequency bins, of said plurality of targets detected on each virtual antenna, to generate the overall set made of said frequency bins that appear at least once in said sub-sets.

(Step 6.2) Method further comprising target energies estimation and fusion thereof, wherein fusion of said target energies includes collecting the energies associated with all the frequency bins of said sub-set to compute the average energy associated with a given frequency bin of said overall set by summing up the square of the absolute value of all the complex amplitudes corresponding to said frequency bin for all said virtual antennas and dividing it by the overall number of antennas that contribute to the computed average energy.

According to another aspect of the present invention, it is herewith disclosed a method, which is per se inventive, for the estimation of the angles here called SPE. It should be clear that the SPE can be implemented independently from the RPE. Indeed, the RPE represents an inventive method for the estimation of ranges, but, in the implementation of the SPE any known method, such as the periodogram method as in "J. Selva, "ML Estimation and Detection of Multiple Frequencies Through Periodogram Estimate Refinement", IEEE Signal Processing Letters, vol. 24, n. 3, pagg. 249-253, mar. 2017" and references therein, for the estimation of the ranges can be implemented to feed the SPE.

In the same manner, RPE can be implemented independently from SPE, being possible to estimate angles from the ranges estimated by RPE, using known techniques, such as 2D FFT processing over the whole receive array as described in "S. Rao, "MIMO Radar," Texas Instruments - Application Report SWRA554A, July 2018, for FMCW radar systems".

Of course, the implementation of both RPE and SPE in succession as disclosed on Figure 5 represents the best solution.

In summary, in the present invention is presented a method for estimating the angular coordinates of point targets whose range has been estimated through a MIMO FMCW radar equipped with a plurality of transmitting (TX) and receiving (RX) antennas, wherein each couple of said TX and RX antennas is replaced with the equivalent virtual antenna; said virtual antennas form a virtual array composed by one or multiple horizontal uniform linear arrays (HULAs) and, optionally, one or multiple vertical uniform linear arrays (VULAs); said MIMO FMCW radar is arranged to generate real or complex signals in response to a propagation scenario including a plurality of point targets; said method includes (see Steps 7 - 11 of Figure 27): (Step 7) acquisition of a spectrum (X₀ ) of said signal and

— (v) its first N_FFT — 1 derivatives ({X_fc ;k = 1,.. N_FFT — 1}) with N_FFT = 3 when said signal is real and N_FFT = 4 when said signal is complex, for each of said virtual antennas,

(Step 8) acquisition of a list of a predetermined number of discretized ranges, named frequency bins, of said previously detected point targets,

(Step 9) selection of a reference antenna, an optional reference vertical uniform linear array, named Reference VULA, a reference horizontal uniform linear array, named Reference HULA, and, optionally, at least a sub-set with a number of said horizontal uniform linear arrays (HULAs) different from the Reference HULA,

(Step 10 [STDAEC]) sequential execution for each of said frequency bins; said execution includes:

- (Step 10.1 [STDAEC-S1]) estimation of the horizontal spatial frequency and complex amplitude of the most dominant point target and optional estimation of the vertical spatial frequency; said estimation includes:

- optional (Step 10.1.1 [STDAE-S1]) computation of the spectrum (^SVULA_,O) °f the sequence collecting the samples of said spectrum (X₀) referred to said Reference VULA in the considered frequency bin and its first 3 derivatives (fs ;k = 1,2,3}) through FFT calculation with an oversampling factor (M),

- optional (Step 10.1.2 [STDAE-S2]) estimation of said vertical spatial frequency, - optional (Step 10.1.3 [STDAE-S3]) combination of the spectra associated with the Reference HULA and said sub-set of HULAs, after compensating for the phase differences between the Reference HULA and said sub-set of HULAs employing said estimate of the vertical spatial frequency ( F_v) ; said superposition, named vertical folding, generates a vertically folded spectrum,

- (Step 10.1.4 [STDAE-S4]) computation of the spectrum {s_{HULA 0} ) and its first three derivatives ({ ^s _{HULA k>} k = 1, 2, 3}) of the sequence collecting the samples of said spectrum (X₀) referred to said Reference HULA in the considered frequency bin, and optionally said spectrum is referred to said vertically folded spectrum if said optional steps (Step 10.1.1 - Step 10.1.3) are carried out; said computations are carried out through FFT calculation with an oversampling factor ( M ) ,

- (Step 10.1.5 [STDAE-S4]) estimation of said horizontal spatial frequency,

- (Step 10.1.6 [STDAE-S5]) checking whether combination of the spectra associated with said Reference HULA, after compensating for their phase differences on the basis of said estimated horizontal frequency and of said estimated vertical frequency, generates an amplitude peak in the resulting spectrum, named overall folded spectrum, wherein the spectra are associated with said virtual antennas of said sub-set of HULAs if said optional steps (Step 10.1.1 - Step 10.1.3) are carried out; and when the checking is positive execution of the following steps:

(Step 10.1.7) refinement of said range and complex amplitude of said most dominant point target;

- optional (Step 10.1.8 [DEEPLEARNING]) compensation for the amplitude distortion due to the non-uniform response of the antenna array for each of said virtual antennas on the basis of said refined complex amplitude (Step 10.1.7), said estimated horizontal frequency and of said estimated vertical frequency employing a deep neural network to compute a new estimate of said complex amplitude for each of said virtual antennas,

(Step 10.2 [STDAEC-S2]) cancellation of said most dominant point target,

- (Step 10.3 [STDAEC-S3]) calculation of a residual energy in the considered frequency bin and its comparison with a predetermined threshold {T_STDAEC), and checking whether said residual energy exceeds said predetermined threshold and, if so, return to said optional computation of the spectrum (Step 10.1.1) if the optional steps are executed or to computation of the spectrum (Step 10.1.4) if the optional steps are not executed, otherwise go to computation of the spatial coordinates (Step 11), otherwise, in negative case of said checking (Step 10.1.6) execution of the following steps:

- (Step 10.4) updating of said acquired list of said frequency bins and return to Step 10 for a further frequency bin or exit when all frequency bins are examined, wherein said sequential execution (Step 10) ends when all the frequency bins have been analyzed, then

(Step 11) computation of the spatial coordinates of all the point targets belonging to all the frequency bins on the basis of said spatial frequencies and said complex amplitude and generation of an overall image of the propagation scenario. Said execution (see Step 10 of Figure 27) for the first embodiment is carried out by running multiple parallel processes whose number is equal to the number of said frequency bins of said acquired list and said generation of the overall image (Step 11 of Figure 27) is carried out when all parallel processes have been executed.

T3) SPE - In this step, each of the L_b frequency bins identified by the RPE are analysed in the angular (i.e., in the azimuth- elevation) domain. This allows to: a) estimate the angular coordinates (i.e., azimuth and elevation) of the target or of the targets whose frequencies (i.e., ranges) fall inside each bin of the set S_b (2.23); b) detect additional targets associated with adjacent frequency bins; c) estimate the angular coordinates (i.e., azimuth and elevation) and compute a finer estimate of the range of such additional targets. It is important to note that many commercial colocated MIMO radar systems are endowed with a ID array (a linear uniform array, usually) and, consequently, can be exploited for 2D imaging only. Our method and its two embodiments can be also used in this case, since the only changes to be made concern the SPE, i.e. T3. In the following, we describe the SPE and how each step is modified to adapt it to 2D imaging. In summary, the above mentioned acquisition (see Step 8 of Figure 27) includes (see Steps 8.1 - 8.2 of Figure 28):

- (Step 8.1) acquisition of a list of a predetermined number of discretized ranges of said plurality of point targets by selecting the indices of distinct frequency bins where at least one point target is detected.

- (Step 8.2) Method further comprising the acquisition of a list of a predetermined number of targets energy corresponding to said discretized ranges.

In this embodiment, the processing accomplished within this task consists of the two consecutive steps listed below (the i-th step is denoted T3-SZ in the following); each step is associated with one of the blocks contained in the SPE block appearing in Figure

7.

T3-S1) Bin analysis - The processing accomplished within this task involves the whole set of triads X-2,v) t v= 0 , 1, N_{VR —} 1} and is executed on a bin-by-bin basis, since it aims at identifying all the targets contributing to the energy of each bin contained in the TRP. For this reason, it consists in running L_b estimators in parallel, one for each of the L_b ranges (i.e., frequency bins) appearing in the TRP (see Figure 7). Moreover, the Z-th estimator executes a novel recursive estimation algorithm, called single target detection, angular estimation and cancellation (STDAEC) and whose description is provided below (additional details can be found in Paragraph 4.1). The STDAEC algorithm processes the N_VH XN_vv matrix

(2.24) collecting the spectral information available on the whole virtual receive array and referring to the ct_j-th frequency bin only (with 1 = 0, 1, L_b — 1), detects L[l] targets and, for each detected target, generates: a) an estimate of its complex amplitude; b) an estimate of its angular coordinates (i.e., its azimuth and its elevation); c) a refined estimate of its range (do not forget that the preliminary estimate of this range is provided by the index a_t of the considered frequency bin). In the following, the azimuth, the elevation and the range characterizing the i-th target detected in the Z-th frequency bin are denoted 0;[Z] , f;[Z] and i?_j[Z], respectively; moreover, its complex amplitude, observed on a reference antenna of the virtual receive array shown in Figure 6 is denoted Q[Z]; in our description of the STDAEC, the selected reference antenna is identified by (p, q) = (J>R, ¾), as shown in Figure 9. In the following, the estimates of the parameters Q;[Z], f;[Z], i?_j[Z] and

Ct[l] computed by the STDAEC are denoted 0;[Z] _r Ri [l] and C;[Z], respectively.

The derivation of the STDAEC algorithm relies on the fact that the Z-th target provides an additive contribution to all the elements of the matrix X[Z] (2.24) and that the phase of this contribution exhibit periodic variations as we move horizontally or vertically along this matrix. More specifically, if we assume that the intensity of the echo received by each RVA from the i- th target does not change from antenna to antenna, the complex amplitude Q[p, q, Z] associated with the Z-th target detected in the Z-th frequency bin and observed on the ( p, q) RVA (i.e., contributing to the spectral coefficient Xi[p, q]; see eqs. (1.7) and (1.10)) where X = c/f₀ is the wavelength associated with the start frequency (see eq. (1.3)) and d_VH ( d_vv ) is the distance between two adjacent horizontal (vertical) RVAs. Consequently, for a given Z, the rate of the phase variations observed in the complex amplitudes {Q[p, q, Z]} as we move horizontally along the considered virtual receive array is proportional to the quantity whereas the rate of the phase variations observed in the same amplitudes as we move vertically along the same array is proportional to

(2.27) The quantities F_{H i}[l] and F_{v i}[l] can be interpreted as the normalized horizontal spatial frequency and the normalized vertical spatial frequency characterizing the i-th target detected in the Z-th frequency bin; if they are known, the elevation and the azimuth of the Z-th target can be evaluated as

(2.28) fi[1] = arcsin and (2.29)

0₍[/] = arcsin respectively. In principle, the horizontal and vertical spatial frequencies can be detected by first computing a 2D DFT of the matrix X[Z] (2.24) and, then, by looking for the local maxima of the absolute values of the elements of the resulting 2D matrix; note that the matrix X[Z]can be also zero-padded before computing its 2D FFT to improve spectral resolution. In the STDAEC algorithm, instead, 2D processing is avoided by alternating vertical and horizontal ID DFT processing; consequently, relevant spatial frequencies are estimated by searching for the peaks of ID amplitude spectra (in other words, an AM approach is adopted). Moreover, in the STDAEC algorithm, the following two additional techniques are exploited: a) Serial target cancellation - This technique is conceptually similar to the cancellation strategy exploited by the STDREC algorithm in T2 and allows to detect multiple targets in the same frequency bin and, in particular, to identify targets having similar spatial coordinates. It is important to point out that the frequencies associated with distinct targets detected in the Z-th frequency bin do not necessarily belongs to that bin; in fact, they can belong to adjacent bins, so that the tails of their spectra are observed in the analysed frequency bin. This problem originates from the fact that, generally speaking, the contribution of a point target to the spectrum computed on each RVA is not a line, unless the associated frequency is exactly a multiple of the DFT frequency /DFT (2.11); consequently, such a contribution is spread over multiple adjacent frequency bins (i.e., spectral tails are observed). b) Spatial folding - As already stated above, the frequency associated with a target detected in the Z-th frequency bin does not necessarily fall exactly in that bin. Folding is employed to get a more accurate estimate of the frequency associated with a given target and is based on the following idea. Once the horizontal and the vertical spatial frequencies associated with a target detected in the Z-th frequency bin have been estimated (see eqs. (2.26) and (2.27)), the spectra {X(p,q)} computed for all the RVAs can be combined in a constructive fashion by

1) taking a reference RVA (identified by (j>, q) = (J>R, ¾); see Figure 9), and compensating for the phase differences, estimated for that target, between the reference RVA and other RVAs, or

2) taking a reference ULA and compensating for the phase differences, estimated for that target, between the reference ULA and other ULAs parallel to the reference ULA.

In case 1), folding generates a single folded spectrum and has the beneficial effects of a) averaging the effects of the noise affecting different RVAs and b) combining in an unconstructive fashion the contributions of all the targets different from the one which the employed spatial frequencies refer to. For this reason, in analysing the amplitude of the folded spectrum, a well defined peak in its amplitude is expected in Z-th frequency bin or in a bin close to it. However, if this does not occur, the detected target is actually a false target (i.e., a ghost target). If a peak exists, its position allows to compute a refined estimate of the frequency (and, consequently, of the range) characterizing the target for which folding has been accomplished. In case 2), folding generates as many folded spectra as the number of antennas of the reference ULA.

In the remaining part of this document, when folding is used, the following terminology is adopted:

Vertical folding - This refers to the case in which folding involves a reference horizontal ULA (HULA) on which other HULAs are folded.

Overall folding - This refers to the case in which folding involves all the spectra {X(p, q)}, i.e. the overall URA (a single folded spectrum is generated in this case).

Note that, in all these cases, the whole virtual receive array or a subset of it can be used. The exploitation of a subset of the available RVAs is motivated by the fact that, in practice, in computing folded spectra, the estimates F_{H i} [l] and F_{v i} [l] of the frequencies F_{H i}[l] and F_{v i}[l] , respectively, are employed and that estimation errors may substantially affect the quality of the phase compensation factors computed for the antennas that are farther from the reference antenna or the reference horizontal/vertical ULA.

The STDAEC algorithm consists of three steps (its r-th step is denoted STDAEC-Sr in the following); the processing accomplished within each of them is sketched below for the Z-th frequency bin (with 1 = 0, 1, ..., L_b — 1). Moreover, the STDAEC algorithm is initialised by

1) Setting the iteration index Z to 1.

2) Setting (2.30)

XM[l] A X[Z].

3) Selecting a reference VULA, consisting of N_VULA adjacent and vertically aligned RVAs (with N_VULA < N_vv), within the virtual receive array; in the following, we assume, without any loss of generality, that the reference VULA includes the reference antenna and, consequently, is identified by q = ¾, q_j + 1, ..., q_F, with q_j < q_R < q_F and P = P_R, as illustrated in Figure 10; note that

4) Selecting a reference HULA, consisting of N_HULA adjacent and horizontally aligned RVAs; in the following, we assume, without any loss of generality, that the reference HULA is the horizontal ULA containing the reference antenna and, consequently, is identified by q = q_R and p = p_j, p_j + 1, ..., p_F, with p_j < p_R < p_F, as illustrated in Figure 10; note that N_HULA = p_F — p_j + 1.

5) Selecting a set of HULAs, different from the reference HULA and having the same size of it; in the following, we assume, without any loss of generality, that these HULAs, called vertically folded HULAs, correspond to

..., q_R — 1, q_R + 1 , ..., , as illustrated in Figure 11; note that the overall number HULAs involved in vertical folding

In summary, the selection (see Step 9 of Figure 27) includes:

- said optional Reference VULA consists of a sub-set of adjacent and vertically aligned virtual antennas,

- said Reference HULA consists of a sub-set of adjacent and horizontally aligned virtual antennas, - each HULA of said optional sub-set of HULAs, different from the reference HULA, has the same size of the reference HULA.

Then, the STDAEC algorithm starts executing its iterations. Within its i-th iteration, it accomplishes the three steps described below.

STDAEC-Sl ) Detection of a new target and estimation of its angular parameters - In this step, the N_VH X N_vv matrix

(2.31) is processed to detect the strongest target contributing to it, and to estimate the angular parameters (0_j[Z],ø;[/]) and the complex amplitude C;[Z] associated with it (this target represents the i- th one detected in the considered frequency bin, since (i— 1) targets have been detected in the previous recursions). The processing accomplished within this step is based on a novel iterative detection and estimation algorithm called single target detection and angular estimation (STDAE) and whose detailed description is provided in Paragraph 4.1; this algorithm consists in the five steps described below (its r-th step is denoted STDAE-Sr in the following). If the STDAE algorithm is used for 2D imaging, the first three steps (i.e., STDAE-S1 to STDAE-S3) have not to be performed; therefore, the first step to be carried out is STDAE-S4. Moreover, the matrix X®[Z] (2.31) is replaced with the N_VH-dimensional vector

(2.32) x⁽⁰M = [*?¾^]]^, collecting the spectral information available on the whole virtual receive array and referring to the Z-th frequency only.

In summary, said estimation (see Step 10.1 of Figure 29) includes a preliminary step (Step 10.1.0) including: a) setting the iteration index i to 1,

— (v) b) defining a matrix collecting said spectral information (X₀ ) available for each of said virtual antennas and referring to the considered frequency bin only, optional c) extracting a vector, referring to said Reference VULA only, from said matrix.

STDAE-S1 ) FFT processing on the reference VULA - The portion of spectral information referring to the reference VULA selected in the initialization is extracted from the matrix X®[Z] and stored in the ZVyy^-dimensional vector

Then, the vector

(2.34) is computed and zero-padded to produce the N₀-dimensional vector with k = 0,1,2 and 3; here,

(2.36) with p = 0, 1, NVULA~ 1, M is a positive integer (dubbed oversampling factor) and

(2.37)

N₀ = M ^■ N_vula.

Finally, the vector

(2.38) LA.fcmHJ is computed for k = 0,1,2 and 3. More specifically, in the following we assume that with m = 0, 1, ..., No— 1.

It is important to point out that: a) Each of the four vectors (^svu_LAfcW^; k = 0, 1, 2, 3} can be computed by executing a iV₀-th order FFT. b) On the one hand, the vector S_yyLA0[Z] collects N₀ equally spaced samples of the spectrum of the sequence _{■■■/■} N_VULA — 1} (see eq. (2.36)). On the other hand, the k-th vector ^svu_LAfcW' with k = 1, 2 and 3, collects, up to a scale factor, N₀ equally spaced samples of the k-th order derivative of the same spectrum.

STDAE-S2) Single angle estimation - The iV₀-dimensional vectors (^svu_LAfcW^; k = 0, 1, 2, 3} are processed to detect the i-th

(strongest target) appearing in Z-th frequency bin and estimate its angular coordinates (this target represents the Z-th one detected in the considered frequency bin, since (Z— 1) targets have been detected in the previous iterations). Note that, generally speaking, the normalised vertical frequency F_{v i}[l] is not a multiple of the fundamental frequency l/iV₀ associated with the FFT processing executed in STDAE-S1 ; for this reason, it can be expressed as where bg_,ί[1\ is an integer belonging to the set {0 , 1 , N₀ 1} and is a real quantity, such that f0 < 5_V:i[l] < 0.5 if bn_,άΆ = 0 (2.41)

] -0.5 < d_n>f[Z]< 0 if b_{n, i}{i\ = N₀ - 1.

(—0.5 < 5_{v i}[l] < 0.5 otherwise

The processing accomplished in this step is executed by a novel estimation algorithm, dubbed complex single frequency estimator (CSFE) and described in detail in Paragraph 4.1.3. This represents the most important part of the STDAE algorithm, since it computes the estimates A_{v i} \l\_r b_nί\1\, [Z] and F_{v i} [l] of the parameters Ai\l\, bn_,ϊ[1\_/ and F_{v i}[l], respectively, on the basis of the set (^svu_LAfcW^; k = 0, 1, 2, 3}. It is worth noting that

Note that the quantity A_{v i} [Z] is not exploited in the following since, it represents a preliminary estimate of A₍[Z].

In summary, the method for the estimation of the vertical spatial frequency (see Step 10.1.2 of Figure 29) can be described as follows (see Steps 10.1.2.1.0 - 10.1.2.1.2 of Figure 30):

- (Step 10.1.2.1.0) representation of a target vertical spatial frequency (F_v), as sum of a main portion (PV^_OFT) ^and ^a residual portion (6_VF_dft), where F_DFT = l/iV₀, where N₀ is the FFT order, wherein said main portion is determined by searching the index (b_n) of the vector element s_VULAOg_v corresponding to the maximum absolute value of the same vector s_VULA0 and said residual frequency computation is optional and can be found through an iterative procedure together with the corresponding target complex amplitude A = aexp(ji|/), where a is said target amplitude and y is said target phase.

(Step 10.1.2.1.1) Said iterative procedure includes a preliminary initialization comprising: setting the iteration index i to 1, the initial estimate of A equal to said vector element (s_VULAOp_v), the initial estimate of D_n = 2TC6_VF_DFT to zero and the initial estimate of F_v to said main portion then iteratively

- (Step 10.1.2.1.2) the i-th iteration is fed by the estimates of D_n, F_v and A computed in the previous iteration, to generate new estimates of D_n, F_v and A, for a predetermined number of iterations, wherein each iteration includes: a) a first step where the residual D_n is calculated solving the first order equation = 0 OR the second order equation where ¾{x} denotes the real part and 3{L:} denotes the imaginary part of the complex quantity x; discarding one of the two solutions, which does not satisfy the inequalities: if b_n = 0 otherwise if said second order equation is solved; then the new estimate of F_v is obtained by adding said residual (A_V F_dft) to said main portion b) a second step wherein said new estimate of F_v is employed to compute the new estimate of A as A = Xj_nt{bn) ^and generated by interpolating the elements of said vector Ms_VULA0 where M is said oversampling factor employed in said FFT calculations. The above mentioned Steps 10.1.2.1.1 - 10.1.2.1.2 of Figure 30 can be replaced by an alternative method including (see Steps 10.1.2.2.1 - 10.1.2.2.2 of Figure 31):

(Step 10.1.2.2.1) Said iterative procedure includes a preliminary initialization comprising: a) setting the iteration index i to 1, the initial estimate of A equal to said vector element (s_VULAOp_v), the initial estimate of D_n = 2TC6_VF_DFT to zero and the initial estimate of F_v to said main portion (PV^DFT); b) computing the real quantity pAF_v/F_DFT; then iteratively

- (Step 10.1.2.2.2) the i-th iteration is fed by the estimates of D_n, F_v, A and the real quantity p computed in the previous iteration, to generate new estimates of D_n, F_v, A and p, for a predetermined number of iterations, wherein each iteration includes: a) a first step where the residual D_n is calculated solving the first order equation = 0 OR the second order equation : where ¾{x} denotes the real part, 3{x} denotes the imaginary part of the complex quantity x and & ⁼ 1,2,3} is computed as: —(VULA) where X_{0 n} is the n-th element of the said sequence collecting the samples of said spectrum (X₀) referred to said Reference VULA in the considered frequency bin; said coefficients can be also computed by interpolating the elements of said spectrum (s_VuLA_,fc)/ discarding one of the two solutions, which does not satisfy the inequalities: if bn — 0 otherwise if said second order equation is solved; then the new estimate of F_v is obtained by adding said residual ( A_V F_dft) to said estimate of F_v ; b) a second step wherein the new estimate of p is computed by dividing said new estimate of F_v by F_dft ; c) a third step wherein said new estimate of F_v is employed to compute the new estimate of A as A = generated by interpolating the elements of said vector M S_VULA,O where M is said oversampling factor employed in said FFT calculations. STDAE-S3) Vertical folding - The estimate F_vi[l\of the normalized vertical frequency F_vi[l\ (2.40) is employed to compensate for the phase difference between each of the HULAs to be folded vertically and the reference HULA (i.e., for the phase differences along the vertical direction), so that all the data of all these HULAs can be combined (i.e., summed) in a constructive fashion. In practice, first, the phase rotation factor

(2.43)

(VF) (VF) _r _,NΊ<7^—

R [i_> q] — exp(-jDy [/])] with

(2.44) is computed for the q-th HULA (with ..., q_R — l, q_R + 1, ..., q^, _f ^VF'⁾)· Then, vertical folding is accomplished by computing the N_HULA ^~dimensional vector 45) that represents a vertically folded spectrum; here r

(2.46) is N_hula-dimensional row vector extracted from the q-th row of the matrix X®[Z] (2.31).

STDAE-S4 ) FFT processing and horizontal frequency estimation - The processing accomplished in this step is very similar to that carried out in STDAE-S1 and STDAE-S2 . In fact, the only difference is represented by the fact that the N_VULA ^~dimensional vector Xvu_LAM (2.33) is replaced by the N_HULA-dimensional vector (2.45) generated in the previous step if the STDAE algorithm is used for 3D imaging and by the vector X®[/] (2.32) if the STDAE algorithm is used for 2D imaging. Therefore, in this case, the CSFE is exploited to estimate the horizontal frequency F_Hi[l] and, again, the complex amplitude i4_j[Z] associated with the Z-th target. Note that, generally speaking, the frequency F_Hi[l] is not a multiple of the fundamental frequency l/iV₀ characterizing the FFT processing executed in STDAE-S1 ; for this reason, it can be expressed

(2.47) where /¾_j[Z] is an integer belonging to the set {0, 1, N₀ — 1} and A_Hi[l] is a real quantity, such that

0< ¾ [Z]< 0.5 f°^rb_H,ί[Z]= 0 (2.48) —0.5< S_Hi[l]< 0 forb_H,i[l]= N₀ -1. -0.5< d_HC[1]£ 0.5 otherwise

The CSFE computes the estimates A_Hi[l], b_Hί[1], d_Hί[1] and F_Hi [Z] of the parameters Ai[1\, /¾_j[Z], <¾_,;[Z] and F_Hi[l] respectively; the estimate F_Hi [Z] is evaluated as

(2.49)

Note that: a) the quantity A_Hi\l\ is not exploited in the following since, it represents a preliminary estimate of Ai\l\; b) the estimates A_vi [Z] and A_Hi\l\ can be significantly different if multiple targets having similar horizontal frequencies or similar vertical frequencies contribute to the considered frequency bin.

In summary, the method for the estimation of the horizontal spatial frequency (see Step 10.1.5 of Figure 29) can be described as follows (see Steps 10.1.5.1.0 - 10.1.5.1.2 of Figure 32):

(Step 10.1.5.1.0) representation of a target horizontal spatial frequency (F_H), as sum of a main portion (bp^rt) and a residual portion (6_HF_dft), where F_DFT = l/iV₀, where N₀ is the FFT order, wherein said main portion is determined by searching the index (b_H) of the vector element s_HULAOg_H corresponding to the maximum absolute value of the same vector s_HULA0 and said residual frequency computation is optional and can be found through an iterative procedure together with the corresponding target complex amplitude A= aexp(ji|/), where a is said target amplitude and y is said target phase.

(Step 10.1.5.1.1) Said iterative procedure includes a preliminary initialization comprising: setting the iteration index i to 1, the initial estimate of A equal to said vector element (s_HULAOp_H), the initial estimate of D_H = 2pd_H F_DFT to zero and the initial estimate of F_H to said main portion ($UF_oft); then iteratively

- (Step 10.1.5.1.2) the i-th iteration is fed by the estimates of D_H, F_H and A computed in the previous iteration, to generate new estimates of D_H, F_H and A, for a predetermined number of iterations, wherein each iteration includes: a) a first step where the residual D_H is calculated solving the first order equation = 0 OR the second order equation where ¾{x} denotes the real part and 3{x} denotes the imaginary part of the complex quantity x; discarding one of the two solutions, which does not satisfy the inequalities: ifb_H =0 otherwise if said second order equation is solved; then the new estimate of F_H is obtained by adding said residual (A_HF_dft) to said main portion (b_H F_dft); b) a second step wherein said new estimate of F_H is employed to compute the new estimate of A as A generated by interpolating the elements of said vector MS_HULAi0 where M is said oversampling factor employed in said FFT calculations.

The above mentioned Steps 10.1.5.1.1 - 10.1.5.1.2 of Figure 32 can be replaced by an alternative method including (see Steps 10.1.5.2.1 - 10.1.5.2.2 of Figure 33):

(Step 10.1.5.2.1) said iterative procedure includes a preliminary initialization comprising: a) setting the iteration index i to 1, the initial estimate of A equal to said vector element (s_HULAOp_H), the initial estimate of D_H = 2pd_H F_DFT to zero and the initial estimate of F_H to said main portion (PH^DFT)/ b) computing the real quantity pAF_H/F_DFT; then iteratively

- (Step 10.1.5.2.2) the i-th iteration is fed by the estimates of D_H, F_h, A and the real quantity p computed in the previous iteration, to generate new estimates of D_H, F_H, A and p, for a predetermined number of iterations, wherein each iteration includes a) a first step where the residual D_H is calculated solving the first order equation = 0 OR the second order equation h: where ¾{x} denotes the real part, denotes the imaginary part of the complex quantity x and ^ ⁼ 1,2,3} is computed as: is the n-th element of the said sequence collecting the samples of said vertically folded spectrum if said optional steps (Step 10.1.1 - Step 10.1.3) are carried out and of the sequence collecting the samples of said spectrum (X₀₎ referred to said Reference HULA in the considered frequency bin if said optional steps are not carried out; said coefficients ({^_HULA,fc_,p<^ ⁼ 1,2,3}) can be also computed by interpolating the elements of said spectrum (s_HULAjfc) ; discarding one of the two solutions, which does not satisfy the inequalities: if b_H = 0 otherwise if said second order equation is solved; then the new estimate of F_H is obtained by adding said residual (A_H F_dft) to said estimate of F_H ; b) a second step wherein the new estimate of p is computed by dividing said new estimate of F_H by F_dft ; c) a third step wherein said new estimate of F_H is employed to compute the new estimate of A as A = generated by interpolating the elements of said vector MS_HULA_,O where M is said oversampling factor employed in said FFT calculations.

STDAE-S5) Overall folding and frequency/amplitude estimation - In this step, the angular information (i.e., the frequencies F_Vti [l] (2.42) and F_{H i} [l] (2.49)) computed in the previous step are exploited to apply folding to the whole receive antenna array or to a portion of it; this step involves the whole spectrum computed on the selected RVAs. If the whole receive antenna array is exploited, overall folding requires the computation of the iV₀-dimensional vectors with m = 0,1,2, where

(2.51) is a phase rotation factor, [l, q] is expressed by eq. (2.43)

(2.52) and note that R_i ^HV'*\l, p, q\ if p = p_R and q = q_R . If the sequence of the absolute value of the elements of the vector X_0,OF[Z] (2.50) is analysed, a relative maximum (i.e., a peak) is expected in the a_j-th frequency bin or in a bin close to it; if this peak is really visible, the SFE algorithm is run to estimate, on the basis of the vectors X_O.OFM _/· XI.OFM and X2_,OFM (see eq. (2.50)), the final estimates

(2.54) and Ai [Z] of the parameters /;[/] and i4_j[Z] characterizing the i-th target detected in the Z-th frequency bin. Note that the integer part ai [Z] of the frequency /i[Z] does not necessarily coincide with (see eq. (2.54)); however, if it differs, it is close to a_t. In the last case, if the quantity ct_j [Z] appears in one of the couples of set S_b (2.23), it is discarded. Otherwise, the new couple

(2.55)

(biUlE_b'Lb), where

(2.56) is added to the set S_b and the number of its elements (denoted L_b) is increased by one. This means that the STDAEC algorithm has to be run on this additional bin too. It is worth noting that if the STDAE algorithm is used for 2D imaging only, the frequency F_vi[l] (2.42) is not available and, therefore, it is not employed in the overall folding.

In summary, the above method can be described as follows: said optional estimate of the vertical spatial frequency (F_v) and said estimate of the horizontal spatial frequency (F_H) are employed to compute the phase difference between said reference antenna and said virtual antennas of said optional sub-set of HULAs to be used in said constructive combination (see Step 10.1.6 of Figure 29) to compute said overall folded spectrum (Xo_F,o) and its first three derivatives ({X_0Ffc;k = 1,2,3}) by combining the spectral contributions referred to said sub-set of virtual antennas. Said checking (see Step 10.1.6 of Figure 29) is performed on the absolute value of said overall folded spectrum

In summary, the method for the refinement (see Step 10.1.6 of Figure 29) can be described as follows (see Steps 10.1.6.1.0 - 10.1.6.1.2 of Figure 34):

(Step 10.1.6.1.0) representation of a target normalized frequency F = f / f_s , where f_s is the sampling frequency, as sum of a main portion (aF_dft) and a residual portion (5F_dft), where F_DFT = l/iV₀, where iV₀ is the FFT order, wherein said main portion is determined by searching the index (a) of the vector element X_OF,o,a corresponding to the element of the vector X_OF,O having the maximum absolute value and said residual frequency is found through an iterative procedure together with the corresponding target complex amplitude A = a exp(jxp), where a is said target amplitude and y is said target phase;

(Step 10.1.6.1.1) Said iterative procedure includes a preliminary initialization comprising: setting the iteration index i to 1, the initial estimate of A equal to said vector element (X_OF,O,¾)_r the initial estimate of A = 2pdF_DFT to zero and the initial estimate of F to said main portion (aF_DFT); then iteratively

- (Step 10.1.6.1.2) the i-th iteration is fed by the estimates of D, F and A computed in the previous iteration, to generate new estimates of D, F and A, for a predetermined number of iterations, wherein each iteration includes a) a first step where the residual D is calculated solving the first order equation Z¾D+ c¾ = 0 OR the second order equation where ¾{x} denotes the real part and 3{x} denotes the imaginary part of the complex quantity x; discarding one of the two solutions, which does not satisfy the inequalities: if a = 0 otherwise if said second order equation is solved; then the new estimate of F is obtained by adding said residual (AF_dft) to said main portion (aF_DFT); b) a second step wherein said new estimate of F is employed to compute the new estimate of A as A = Xi_nt{P) is generated by interpolating the elements of said vector MX_OFO where M is said oversampling factor employed in said FFT calculations.

The above mentioned Steps 10.1.6.1.1 - 10.1.6.1.2 of Figure 34 can be replaced by an alternative method including (see Steps 10.1.6.2.1 - 10.1.6.2.2 of Figure 35):

(Step 10.1.6.2.1) Said iterative procedure includes a preliminary initialization comprising: a) setting the iteration index i to 1, the initial estimate of A equal to said vector element (X_OF_,o_,a)_r the initial estimate of A = 2pd F_DFT to zero and the initial estimate of F to said main portion (aF_DFT); b) computing the real quantity p = F/F_DFT; then iteratively

- (Step 10.1.6.2.2) the i-th iteration is fed by the estimates of D, F and A and the real quantity p computed in the previous iteration, to generate new estimates of D, F, A and p, for a predetermined number of iterations, wherein each iteration includes: a) a first step where the residual D is calculated solving the first order equation b^A + C_Q = 0 OR the second order equation ¾D² + ¾D + ¾ = 0 in the variable D with: where ¾{x} denotes the real part, 3{x} denotes the imaginary part of the complex quantity x and {Xo_F,k,p>k = 1,2,3} is computed by interpolating the elements of said overall folded spectrum

(Xo_F,fc); discarding one of the two solutions, which does not satisfy the inequalities: if a = 0 otherwise if said second order equation is solved; then the new estimate of F is obtained by adding said residual (AF_dft) to said estimate of F; b) a second step wherein the new estimate of p is computed by dividing said new estimate of F by F_dft; c) a third step wherein said new estimate of F is employed to compute the new estimate of A as is generated by interpolating the elements of said vector MX_OFO where M is said oversampling factor employed in said FFT calculations.

STDAEC-S2) Target cancellation - The contribution C^[Z], given by the i-th target detected in the Z-th frequency bin, to the vector X®[Z] (2.31) is computed (see eqs. (4.230)-(4.231)) and is cancelled. Cancellation consists in the computation of the new residual vector

(2.57)

In summary, the above method can be described as follows: said cancellation (see Step 10.2 of Figure 29) includes the computation and subtraction of the contribution given by said dominant point target to the spectral contribution referred to said considered frequency bin to generate a new residual of the spectrum and of the corresponding derivatives.

STDAEC-S3) Residual energy test - The energy of the residual spectrum vector X^ⁱ⁺¹^[Z] (2.57) is compared with the positive threshold ^STDAEC (which may exhibit an angle dependence). If this energy is below the threshold, the STDAEC algorithm stops; otherwise, the recursion index i is increased by one and a new iteration is started by going back to STDAEC- S1.

If D[l] iterations are accomplished by the STDAEC algorithm operating on the a_j-th frequency bin, no more than D[l] distinct targets are identified in that bin (D[Z] targets are found if none of them is deemed to be a ghost target). All the target information acquired from the a_j-th frequency bin are collected in the set

(2.59) 0,1. D[l]- l}, with 1= 0, 1, ..., L_{b —} 1. Note that, as already mentioned for the

STDREC algorithm, the STDAEC algorithm can be easily extended in a way that multiple angles are detected and estimated in parallel. This means that multiple instances of the CSFE algorithm are run in parallel.

T3-S2) Evaluation of spatial coordinates and generation of the overall image - This task aims at: a) computing the spatial coordinates of all the targets on the basis of the spatial information generated by the L_b STDAEC algorithms and collected in the L_b sets {?]} (see eq. (2.59)); b) generating the overall image of the propagation scenario. More specifically, the estimates of the range, of the elevation and of the azimuth of the i-th target detected in the a_j-th frequency bin are computed as (see eqs. (1.7)-(1.9) and (2.28)-(2.29))

(2.60)

<Pi [*]= arcsin and

9i [l] = arcsin respectively. Finally, these information are fused to generate the overall set

(2.63) 0,1. L — l}, describing the generated radar image, which, generally speaking, is a cloud of L points. The set J_t results from the union of all the sets {J®}, where

(2.64) 0,1. D[l] - 1], with 1 = 0, 1, ..., L_b — 1 .

It is worth stressing that if the STDAE algorithm is used for 2D imaging only, the frequency F_{v i} [l\ (2.42) is not available and, therefore, it is not included in the set J] (2.59) and, as a consequence, the elevation angle f.[Z] (2.61) can not be estimated.

In summary, the method for the evaluation of the spatial coordinates can be summarized as follows: if the optional steps (Step 10.1.1 - Step 10.1.3 of Figure 29) are carried out, said computation (see Step 11 of Figure 27) includes the computation of the range (R) , the elevation (f ) and the azimuth (Q) as: f = arcsin and q = arcsin respectively, and, if the optional steps (Step 10.1.1 - Step 10.1.3 of Figure 29) are not carried out, said computation (see

Step 11 of Figure 27) includes the computation of the range (R) and the azimuth (Q) as: and

Q = arcsin respectively, here, m is the radar chirp rate, c is the speed of light, l is the radar wavelength, d_vv is the distance between two adjacent virtual antennas of said VULAs and d_VH is the distance between two adjacent virtual antennas of said HULAs

2 .2 .2 . Embodiment #2 The processing accomplished in the second embodiment of the proposed method is described by the block diagram illustrated in Figure 12 and can be divided in three tasks, each associated with one of the blocks appearing in that figure (the i-th task is denoted Ti in the following). In all these tasks, a single array snapshot is processed. The first two tasks of this embodiment (namely, those concerning FFT processing and RPE) coincide with that of the first embodiment. For this reason, in the remaining part of this paragraph, the third task only is sketched; additional details about the algorithms employed in this task are provided in Paragraph 4.2.

In the third task of this embodiment, spatial estimation is accomplished serially i.e., on a bin-by-bin basis; moreover, the identified frequency bins are analysed in an ordered way and, more precisely, according to decreasing energies; in the following we assume, for simplicity, that the bins in the set S_b (2.23) have been ordered according to decreasing energies, so that

(2.65)

E_b,i ³ E_{b i-l} for l= 0,1,...,L_{b —} 1.

In summary, for the second embodiment, said acquisition (see Step 8 of Figure 27) includes the acquisition of target energy and wherein said execution (see Step 10 of Figure 27) is repeated sequentially over the acquired list, ordered according to said target energy, and wherein said generation of the overall image (Step 11 of Figure 27) is carried out when all the frequency bins have been analyzed.

The SPE is initialised by setting

(2.66) g(°) g

^th,n — ^m,v> with m = 0,l,2 and the bin index 1 to 0. Then, L_b consecutive iterations are run. In the Z-th iteration, the STDAEC algorithm is run to a) estimate, on the basis of the N_VH X N_vv matrix

(2.67) the angular parameters of the D[l] distinct targets contributing to that bin. A refined estimate of the range for all the targets detected in the considered frequency bin is computed on the basis of set of triads {(X® , X® , X®)}. Then, the estimates of the range and of the angular parameters of the D[l] detected targets are: a) made available at the output of the SPE; b) processed by another algorithm, called single bin cancellation (SBC). The SBC algorithm estimates the contributions C®_v)} of these targets (i.e., of the Z-th frequency bin) to the set of triads

{(X® , X® , X®)} (see eqs. (4.11)-(4.13)) and cancels it. Cancellation consists in the computation of the new residual triad (2.68) with m = 0,l,2 and v 0, 1, ..., N_VR 1. Finally, if l L_b 1, l is increased by one and a new iteration is started; otherwise, if l = L_b 1, processing is over. A similar adaptation for 2D imaging can be derived also for the second embodiment since it is based on the same algorithms adopted in the first one.

In summary, the above method can be described as follows: said cancellation (see Step 10.2 of Figure 29) includes the computation and subtraction of the contribution given by said

— (v) dominant point target to said spectrum (X₀ ) and to the corresponding derivatives to generate a new residual of the spectrum and of the corresponding derivatives.

2.3. SFCW radar system

The two embodiments described above for a FMCW radar system can be easily adapted to the case of a SFCW radar system.

In summary, in the present invention is presented a method for the detection of targets through a MIMO SFCW radar equipped with a plurality of transmitting (TX) and receiving (RX) antennas, wherein each couple of said TX and RX antennas is replaced with the equivalent virtual antenna; said MIMO SFCW radar is arranged to generate complex signals in response to a propagation scenario including a plurality of point targets; said method includes (see Steps 1 - 6 of Figure 36): - (Step 1: IFFT processing) estimation of the impulse response

— (v)

(X₀ ) characterizing the communication channel associated with the said equivalent virtual antenna and its first N_FFT — 1

— (v) derivatives ({X_fc ;k = 1,.. N_FFT — 1}) with N_FFT = 4, for each of said virtual antennas through IFFT calculation, with an oversampling factor (M),

- (Step 2: RVA selection) acquisition of a sub-set of a number (N_a) of virtual antennas impulse response and of the corresponding derivatives,

Recursive execution of

(Step 3: STDREC-S1) calculation of an estimate of the parameters of the most dominant point target, including phase (y), amplitude (a) and delay (t) for each of said virtual antennas of said sub-set on the basis of said impulse response, and its derivatives, through N_FFT — 1 IFFTs for each of said virtual antennas of said sub-set,

- (Step 4: STDREC-S2) cancellation of said most dominant point target and computation of a residual impulse response,

- (Step 5: STDREC-S3) computation of an energy of said residual impulse response and return to (Step 3) until said energy is over a predetermined threshold (T_STDREC)_r otherwise

- (Step 6) computation of an average energy and fusion of ranges information of said point targets detected by said number of virtual antennas to list said point targets according to said average energy in a decreasing order. 2.3.1. Embodiment #1

In this paragraph we analyse the main changes that need to be introduced in the embodiment #1 to adapt it to a MIMO SFCW radar system. As far as T1 is concerned, the only required change is represented by the replacement of: a) the vector r_v (1.6) with the vector x_v (1.22); b) each FFT with an inverse FFT (IFFT). This means that the spectra {X₀ Xi _{,V r} Xz_{,v /} v = 0, 1, N_{VR —} 1} computed in T1 (now called IFFT processing) provide time domain information and, in particular, information about the impulse responses of the communication channels associated with all the available RVAs. For this reason, frequency bins are replaced by bins in the time domain; such bins are called delay bins in the remaining part of this document. The changes to be made in T2 concern its second step (i.e., the

STDREC algorithm), since the noisy measurements processed in this step are always complex (see eqs. (1.22)-(1.23)); the new structure of this step is summarized below.

T2-S2) Target detection and range estimation - In this step, that aims at detecting the most relevant targets on each of the N_A antennas, a minor change is required in the cancellation procedure with respect to its counterpart employed in the MIMO FMCW radar system. This is due to the fact that, as already stated above, the noisy measurements processed in a MIMO SFCW radar system are always represented by complex (frequency- domain) sequences.

The initialization of the STDREC algorithm remains unchanged: the vectors are defined on the basis of eq.

(2.10) (with k = 0, 1, ..., N_A — 1 ) . In its i-th iteration, this algorithm accomplishes the three steps described below (these are denoted STDREC-Sp in the following, with p = 1,2 and 3).

STDREC-S1 ) Detection of a new target and estimation of its parameters - In this step, the delay t® and the complex amplitude associated with the i-th target are estimated. Note that, generally speaking, the delay is not a multiple of of the fundamental delay characterizing the output of the IFFT processing executed in task Tl ; for this reason, it can be expressed as

(2.70) t (0 _ (0 , r(i) v_k - ^av_k ^TIDFT + S_Vk ^TIDFT< where a® is an integer parameter belonging to the set {0, 1, ...,

N₀ — 1} and 5® is ^a real parameter, such that the conditions expressed by eq. (2.13) are satisfied. The processing accomplished in this step is executed by an algorithm dubbed complex single delay estimator (CSDE) in the following (see Paragraph 4.1.4); this computes the estimates

T¾ = ^avl ^TIDFT + S_Vk ^TIDFT of the parameters /i®, respectively, on the basis of the triad (X¾_k, X¾_fe, ^XS_fe)·

In summary, the above method can be described as follows: when a base-band signal made available by said radar is complex with in phase and quadrature components, Step 3 of Figure 36 includes (see Steps 3.5.0 - 3.5.2 of Figure 37):

- (Step 3.5.0) representation of a target normalized delay T = tD/, where D/ is the frequency step size of the employed radar, as sum of a main portion (QCF_idft) and a residual portion (_<5F_idft), where F_IDFT = l/iV₀, where N₀ is the IFFT order, wherein said main portion is determined by searching the index (a) of the vector

— (v) element X_oa corresponding to the maximum absolute value of the

— (v) same vector X₀ and said residual portion is found through an iterative procedure together with the corresponding target complex amplitude A = aexp(-jxp), where a is said target amplitude and y is said target phase.

- (Step 3.5.1) Said iterative procedure includes a preliminary initialization comprising: setting the iteration index i to 1, the initial estimate of A (v) equal to said vector element (X_0,a)_r the initial estimate of D = 2i F_IDFT to zero and the initial estimate of T to said main portion (aF_IDFT); then iteratively

- (Step 3.5.2) the i-th iteration is fed by the estimates of D, T and A computed in the previous iteration, to generate new estimates of D, T and A, for a predetermined number of iterations, wherein each iteration includes a) a first step where the residual D is calculated solving the first order equation Z¾D + c¾ = 0 OR the second order equation a_dD² + Z¾D + C_Q = 0 in the variable D with: where ¾{x} denotes the real part and 3{x} denotes the imaginary part of the complex quantity x; discarding one of the two solutions, which does not satisfy the inequalities: if a = 0 otherwise if said second order equation is solved; then the new estimate of T is obtained by adding said residual (A F_idft) to said main portion (3 F_idft); b) a second step wherein said new estimate of T is employed to compute the new estimate of A as A = Xi_nt(f) and Xin_tij^ is generated

— O) by interpolating the elements of said vector MX₀ where M is said oversampling factor employed in said IFFT calculation. The above mentioned Steps 3.5.1 and 3.5.2 of Figure 37 can be replaced by an alternative method including (see Steps 3.6.1 - 3.6.2 of Figure 38):

- (Step 3.6.1) Said iterative procedure includes a preliminary initialization comprising: a) setting the iteration index i to 1, the initial estimate of A

—(v) equal to said vector element the initial estimate of D =

2i F_IDFT to zero and the initial estimate of T to said main portion (a F_IDFT) ; b) computing the real quantity p = T/F_IDFT; then iteratively

- (Step 3.6.2) the i-th iteration is fed by the estimates of D, T, A and the real quantity p computed in the previous iteration, to generate new estimates of D, T, A and p, for a predetermined number of iterations, wherein each iteration includes: a) a first step where the residual D is calculated solving the first order equation b_%A + <¾ = 0 OR the second order equation ¾D² + ¾D + ¾ = 0 in the variable D with: where ¾{x} denotes the real part, 3{L:} denotes the imaginary part

—(v) of the complex quantity x and {X_k,p>k = 1,2,3} is computed as:

(_V) where x_cn is the n-th element of the said complex base-band

—(v) signal acquired on the v-th antenna; said coefficients ({¾ _,p>'k ⁼ 1,2,3}) can be also computed by interpolating the samples of said

— (v) impulse response (X_¾ ); discarding one of the two solutions, which does not satisfy the inequalities: ifa = 0 otherwise if said second order equation is solved; then the new estimate of T is obtained by adding said residual (A F_idft) to said estimate of T; b) a second step wherein the new estimate of p is computed by dividing said new estimate of T by F_idft ; c) a third step wherein said new estimate of T is employed to compute the new estimate of A as A = Xi_nt(f) and Xi_nt(T) is generated

— O) by interpolating the elements of said vector MX₀ where M is said oversampling factor employed in said IFFT calculations.

STDREC-S2) Cancellation of the new target - The contributions by the i-th (i.e., by the last) target detected on the v_k-th RVA to the triad (X_g^_fc, , X®_fe), are computed on the basis of eqs. (4.214)-(4.216) (see Paragraph

4.1.6). Therefore, the cancellation procedure consists again in the computation of the new residual triad on the basis of eq. (2.15).

In summary, the above method can be summarized by considering that said cancellation, see Step 4 of Figure 36, includes the computation and subtraction of the contribution given by said

— (v) dominant target to said impulse response (X₀ ) and to the corresponding derivatives to compute a residual of the impulse response and of the corresponding derivatives. STDREC-S3) Computation of the residual energy in the time domain - The energy of the residual time-domain vector is computed and compared with the positive threshold TSTDREC. If this energy is below the threshold, the STDREC algorithm stops and L_k = i relevant targets are detected on the v_k-th RVA; otherwise, the recursion index i is increased by one and a new recursion is started (i.e., we go back to STDREC-S1). Moreover, the following interpretation for the processing accomplished by the STDREC on the v_k-th RVA can be given. The vector X_0Vfe ^can be seen as a collection of noisy information referring to N₀ distinct delay bins (i.e., to N₀ distinct range bins) and is usually dense in the presence of multiple extended targets. The STDREC allows to extract a discrete delay (i.e., range) profile from the vector X_0Vfe; such a profile is usually sparse.

T2-S3) Range information fusion - In this case, the set _<A_b (see eq. (2.17)) collects L_b relevant time bins {a_t; 1= 0, 1, ..., L_b —

1}, whereas the set S_b contains the associated energies (see eq. (2.18)). The average energy associated with the a₍-th bin (with 1= 0, 1, ..., L_b — 1 ) is where N_bi is defined by eq. (2.22). The final output of the information fusion is represented by the set S_b (2.23). Note that the cardinality L_b of the last set represents a preliminary estimate of the overall number of targets.

In summary, the above fusion step (see Step 6 of Figure 36) can be described as follows (see Steps 6.1 - 6.2 of Figure 26):

- (Step 6.1) fusion of the information provided by a number (N_A) of sub-sets collecting all the discretized ranges, named delay bins, of said plurality of targets detected on each virtual antenna, to generate the overall set made of said delay bins that appear at least once in said sub-sets.

(Step 6.2) Method further comprising targets energies estimation and fusion thereof, wherein fusion of said target energies includes collecting the energies associated with all the delay bins of said sub-sets to compute the average energy associated with a given delay bin of said overall set by summing up the square of the absolute value of all the complex amplitudes corresponding to said delay bin for all said virtual antennas and dividing it by the overall number of antennas that contribute to the computed average energy.

According to another aspect of the present invention, it is herewith disclosed a method, which is per se inventive, for the estimation of the angles here called SPE.

It should be clear that the SPE can be implemented independently from the RPE. Indeed, the RPE represents an inventive method for the estimation of ranges, but, in the implementation of the SPE any known method, such as the periodogram method, for the estimation of the ranges can be implemented to feed the SPE. In the same manner, RPE can be implemented independently from SPE, being possible to estimate angles from the ranges estimated by RPE, using known techniques, such as 2D IFFT processing over the whole receive array.

Of course, the implementation of both RPE and SPE in succession as disclosed on Figure 5 represents the best solution.

In summary, in the present invention is presented a method for estimating the angular coordinates of point targets whose range has been estimated through a MIMO SFCW radar equipped with a plurality of transmitting (TX) and receiving (RX) antennas, wherein each couple of said TX and RX antennas is replaced with the equivalent virtual antenna; said virtual antennas form a virtual array composed by one or multiple horizontal uniform linear arrays (HULAs) and, optionally, one or multiple vertical uniform linear arrays (VULAs); said MIMO SFCW radar is arranged to generate complex signals in response to a propagation scenario including a plurality of point targets; said method includes (see Steps 7 - 11 of Figure 27):

— (v)

(Step 7) acquisition of an impulse response (X₀ ) characterizing the communication channel associated with the said equivalent virtual antenna and its first N_FFT — 1 derivatives

— (v)

({X_fc ; k = 1,... , N_FFT — 1}) with N_FFT = 4, for each of said virtual antennas,

(Step 8) acquisition of a list of a predetermined number of discretized ranges, named delay bins, of said previously detected point targets,

(Step 9) selection of a reference antenna, an optional reference vertical uniform linear array, named Reference VULA, a reference horizontal uniform linear array, named Reference HULA, and, optionally, at least a sub-set with a number of said horizontal uniform linear arrays (HULAs) different from the Reference HULA,

- (Step 10 [STDAEC]) sequential execution for each of said delay bins; said execution includes:

- (Step 10.1 [STDAEC-S1]) estimation of the horizontal spatial delay and complex amplitude of the most dominant point target and optional estimation of the vertical spatial delay; said estimation includes:

- optional (Step 10.1.1 [STDAE-S1]) computation of the spectrum (^SVULA_,O) °f the sequence collecting the samples of said impulse response (X₀) referred to said Reference VULA in the considered delay bin and its first 3 derivatives ({^s _{VULA k>} k = 1, 2, 3}) through IFFT calculation with an oversampling factor ( M ) ,

- optional (Step 10.1.2 [STDAE-S2]) estimation of said vertical spatial delay,

- optional (Step 10.1.3 [STDAE-S3]) combination of the spectra associated with the Reference HULA and said sub-set of HULAs, after compensating for the phase differences between the Reference HULA and said sub-set of HULAs employing said estimate of the vertical spatial delay ( T_v ) ; said superposition, named vertical folding, generates a vertically folded impulse response,

- (Step 10.1.4 [STDAE-S4]) computation of the spectrum {s_{HULA 0} ) and its first three derivatives ({ ^s _{HULA k>} k = 1, 2, 3}) of the sequence collecting the samples of said impulse response (X₀) referred to said Reference HULA in the considered delay bin, and optionally said spectrum is referred to said vertically folded impulse response if said optional steps (Step 10.1.1 - Step 10.1.3) are carried out; said computations are carried out through IFFT calculation with an oversampling factor (M),

- (Step 10.1.5 [STDAE-S4]) estimation of said horizontal spatial delay,

- (Step 10.1.6 [STDAE-S5]) checking whether combination of the impulse responses associated with said virtual antennas of said Reference HULA, after compensating for their phase differences on the basis of said estimated horizontal delay and of said estimated vertical delay, generates an amplitude peak in the resulting impulse response, named overall folded impulse response, wherein the impulse responses are associated with said virtual antennas of said sub-set of HULAs if said optional steps (Step 10.1.1 - Step 10.1.3) are carried out; and when the checking is positive execution of the following steps:

(Step 10.1.7) refinement of said range and complex amplitude of said most dominant point target;

- optional (Step 10.1.8 [DEEPLEARNING]) compensation for the amplitude distortion due to the non-uniform response of the antenna array for each of said virtual antennas on the basis of said refined complex amplitude (Step 10.1.7), said estimated horizontal delay and of said optionally estimated vertical delay employing a deep neural network to compute a new estimate of said complex amplitude for each of said virtual antennas,

(Step 10.2 [STDAEC-S2]) cancellation of said most dominant point target, - (Step 10.3 [STDAEC-S3]) calculation of a residual energy in the considered delay bin and its comparison with a predetermined threshold {T_STDAEC), and checking whether said residual energy exceeds said predetermined threshold and, if so, return to said optional computation of the spectrum (Step 10.1.1) if the optional steps are executed or to computation of the spectrum (Step 10.1.4) if the optional steps are not executed, otherwise go to computation of the spatial coordinates (Step 11) otherwise, in negative case of said checking (Step 10.1.6) execution of the following steps:

- (Step 10.4) updating of said acquired list of said delay bins and return to Step 10 for a further delay bin or exit when all delay bins are examined, wherein said execution (Step 10) ends when all the delay bins have been analyzed, then

(Step 11) computation of the spatial coordinates of all the point targets belonging to all the delay bins on the basis of said spatial delays and said complex amplitude and generation of an overall image of the propagation scenario.

Said execution (see Step 10 of Figure 27) for the first embodiment is carried out by running multiple parallel processes whose number is equal to the number of said delay bins of said acquired list and said generation of the overall image (Step 11 of Figure 27) is carried out when all parallel processes have been executed.

The processing accomplished in T3 (i.e., by the SPE) has the same structure and interpretation as that illustrated for the FMCW radar system since it aims at estimating the angular parameters of the targets on the basis of the discrete range profile computed in T2. However, frequency bins are now replaced by delay bins and, consequently, the normalised horizontal and vertical frequencies F_{H i} and F_{v i} by the normalised horizontal and vertical delays T_{H i} and T_{v i r} respectively. Note also that, in step T3-S1, the matrix X[Z] is still defined on the basis of eq. (2.24), but the complex amplitude Ai[p, q, l\ appearing in eq. (2.25) is replaced by (see eqs. (1.7) and (1.23)-(1.24)): where i4_j[Z] is the complex amplitude observed for the i-th target in the Z-th time bin on the reference RVA. This equation leads to some important differences in the formulas employed by the STDAE algorithm in the estimation of the parameters (qi[1], fi[1],T_j[Z],i4₍[Z]) (see STDAEC-S1) with respect to the case of a MIMO FMCW radar system. In summary, the above mentioned acquisition (see Step 8 of Figure 27) includes (see Steps 8.1 - 8.2 of Figure 28):

- (Step 8.1) acquisition of a list of a predetermined number of discretized ranges of said plurality of targets by selecting the indices of distinct delay bins where at least one point target is detected.

- (Step 8.2) Method further comprising the acquisition of a list of a predetermined number of targets energy corresponding to said discretized ranges.

In summary, similarly as the FMCW radar system, the selection (see Step 9 of Figure 27) includes: - said optional Reference VULA consists of a sub-set of adjacent and vertically aligned virtual antennas,

- said Reference HULA consists of a sub-set of adjacent and horizontally aligned virtual antennas, - each HULA of said optional sub-set of HULAs, different from the reference HULA, has the same size of the reference HULA.

In summary, similarly as the FMCW radar system, said estimation (see Step 10.1 of Figure 29) includes a preliminary step (Step 10.1.0) including: a) setting the iteration index i to 1, b) defining a matrix collecting the samples of the said impulse

— (v) response (X₀ ) available for each of said virtual antennas and corresponding to the considered delay bin only, optional c) extracting a vector, referring to said Reference VULA only, from said matrix.

STDAE-S1 ) IFFT processing on the reference VULA - In this step, the vector

(2.75) is evaluated for k = 0,1,2 and 3; note that with m = 0, 1, N₀ — 1, and that the quantity is defined by eq. (2.34).

STDAE-S2) Single angle estimation - A representation similar to that expressed by eq. (2.40) for the vertical frequency F_vi[l\ can be adopted for normalised vertical delay T_Vi[l]; in this case, the quantities bg_,ί[1] and A_Fj[Z] are computed by the CSDE algorithm (see Paragraph 4.1.4).

In summary, the method for the estimation of the vertical spatial frequency (see Step 10.1.2 of Figure 29) can be described as follows (see Steps 10.1.2.1.0 - 10.1.2.1.2 of Figure 39):

- (Step 10.1.2.1.0) representation of a target vertical spatial delay (T_v), as sum of a main portion (PV^IDFT) and a residual portion (6_VF_idft), where F_IDFT = l/iV₀, where N₀ is the IFFT order, wherein said main portion is determined by searching the index (b_n) of the vector element s_VULAOg_v corresponding to the maximum absolute value of the elements of the vector S_VULA,O and said residual delay computation is optional and can be found through an iterative procedure together with the corresponding target complex amplitude A= aexp(—ji|/), where a is said target amplitude and y is said target phase.

(Step 10.1.2.1.1) Said iterative procedure includes a preliminary initialization comprising: setting the iteration index i to 1, the initial estimate of A equal to said vector element (s_Vu_LA,o_,j3v)' the initial estimate of D_n A 2TC6_VF_IDFT to zero and the initial estimate of T_v to said main portion (PV^IDFT); then iteratively - (Step 10.1.2.1.2) the i-th iteration is fed by the estimates of D_n, T_v and A computed in the previous iteration, to generate new estimates of D_n, T_v and A, for a predetermined number of iterations, wherein each iteration includes: a) a first step where the residual D_n is calculated solving the first order equation = 0 OR the second order equation a_b-Dg+B_bDg+ c-_b = 0 in the variable D_n with: where ¾{x} denotes the real part and 3{L:} denotes the imaginary part of the complex quantity x; discarding one of the two solutions, which does not satisfy the inequalities: ίίb_n = 0 otherwise if said second order equation is solved; then the new estimate of T_v is obtained by adding said residual (A_V F_idft ) to said main portion (PV^IDFT)/ b) a second step wherein said new estimate of T_v is employed to compute the new estimate of A as A = generated by interpolating the elements of said vector M S_VULA_,O where M is said oversampling factor employed in said IFFT calculations. The above mentioned Steps 10.1.2.1.1 - 10.1.2.1.2 of Figure 39 can be replaced by an alternative method including (see Steps 10.1.2.2.1 - 10.1.2.2.2 of Figure 40):

(Step 10.1.2.2.1) said iterative procedure includes a preliminary initialization (Step 10.1.2.2.1) comprising: a) setting the iteration index i to 1, the initial estimate of A equal to said vector element (s_VULAOp_v), the initial estimate of D_n = 2TC6_VF_IDFT to zero and the initial estimate of T_v to said main portion (PV^IDFT); b) computing the real quantity p A T_v/F_IDFT ; then iteratively

- (Step 10.1.2.2.2) the i-th iteration is fed by the estimates of D_n, T_v, A and the real quantity p computed in the previous iteration, to generate new estimates of D_n, T_v, A and p, for a predetermined number of iterations, wherein each iteration includes a) a first step where the residual D_n is calculated solving the first order equation = 0 OR the second order equation where ¾{x} denotes the real part, 3{x} denotes the imaginary part of the complex quantity x and & ⁼ 1,2,3} is computed as: where is the n-th element of the said sequence collecting the samples of said impulse response referred to said Reference VULA in the considered frequency bin; said coefficients ({svuLA.fc.pi& ⁼ 1,2,3}) can be also computed by interpolating the samples of said spectrum (S_VULA_,/C)/ discarding one of the two solutions, which does not satisfy the inequalities: if said second order equation is solved; then the new estimate of T_v is obtained by adding said residual (A_VF_idft) to said estimate of T_v; b) a second step wherein the new estimate of p is computed by dividing said new estimate of T_v by F_idft; c) a third step wherein said new estimate of T_v is employed to compute the new estimate of A as A = Xi_nt(^'f_v) and Xi_nt generated by interpolating the elements of said vector Ms_VULA0 where M is said oversampling factor employed in said IFFT calculations.

STDAE-S3) Vertical folding - Vertical folding is employed to compensate for the phase differences between the considered HULAs. However, the phase rotation factor appearing in eq. (2.45) is computed as (2.77) where and T_{V i} [l] is an estimate of the normalized vertical delay T_{V i}[l\ (evaluated in STDAE-S2 ) .

STDAE-S4 ) IFFT processing and horizontal delay estimation - In this step, the CSDE algorithm is exploited to compute the estimates T_{H i} [l] and Ai [Z] of the normalized horizontal delay T_{H i}[l] and of the complex amplitude Fj[Z], respectively; both are associated with the i-th target. A similar representation as that expressed by eq. (2.47) can be adopted for the delay T_{H i}[l\ ; the quantities /¾_j[Z] and A_{H i}[l] are estimated by the CSDE (see

Paragraph 4.1.4).

In summary, the method for the estimation of the horizontal spatial frequency (see Step 10.1.5 of Figure 29) can be described as follows (see Steps 10.1.5.1.0 - 10.1.5.1.2 of Figure 41): - (Step 10.1.5.1.0) representation of a target horizontal spatial delay (T_H), as sum of a main portion (P_H^_IDFT) and a residual portion (6_HF_idft), where F_IDFT = l/iV₀, where N₀ is the IFFT order, wherein said main portion is determined by searching the index (b_H) of the vector element s_HULAOg_H corresponding to the maximum absolute value of the elements of the vector S_HULA,O and said residual delay computation is optional and can be found through an iterative procedure together with the corresponding target complex amplitude A = a exp(— ji|/) , where a is said target amplitude and y is said target phase.

(Step 10.1.5.1.1) Said iterative procedure includes a preliminary initialization comprising: setting the iteration index i to 1, the initial estimate of A equal to said vector element (¾_ULAi0j3_H) _r the initial estimate of D_H = 2pd_H F_IDFT to zero and the initial estimate of T_H to said main portion (PH^IDFT); then iteratively - (Step 10.1.5.1.2) the i-th iteration is fed by the estimates of D_H , T_H and A computed in the previous iteration, to generate new estimates of D_H , G_H and A, for a predetermined number of iterations, wherein each iteration includes: a) a first step where the residual D_H is calculated solving the first order equation + Cp = 0 OR the second order equation where ¾{x} denotes the real part and 3{x} denotes the imaginary part of the complex quantity x; discarding one of the two solutions, which does not satisfy the inequalities: if PH ⁼ 0 otherwise if said second order equation is solved; then the new estimate of T_H is obtained by adding said residual (A_HF_idft) to said main portion (PH^IDFT); b) a second step wherein said new estimate of T_H is employed to compute the new estimate of A as A generated by interpolating the elements of said vector Ms_HULA0 where M is said oversampling factor employed in said IFFT calculations.

The above mentioned Steps 10.1.5.1.1 - 10.1.5.1.2 of Figure 41 can be replaced by an alternative method including (see Steps 10.1.5.2.1 - 10.1.5.2.2 of Figure 42):

(Step 10.1.5.2.1) said iterative procedure includes a preliminary initialization including: a) setting the iteration index i to 1, the initial estimate of A equal to said vector element (s_HULAOp_H), the initial estimate of

D_H = 2pd_H F_IDFT to zero and the initial estimate of T_H to said main portion (PH^IDFT)/ b) computing the real quantity pAF_H/F_IDFT; then iteratively - (Step 10.1.5.2.2) the i-th iteration is fed by the estimates of D_H, G_H, A and the real quantity p computed in the previous iteration, to generate new estimates of D_H, G_H, A and p, for a predetermined number of iterations, wherein each iteration includes a) a first step where the residual D_H is calculated solving the first order equation = 0 OR the second order equation where ¾{x} denotes the real part, denotes the imaginary part of the complex quantity x and is computed as:

—(HULA) where X_{0 n} is the n-th element of the said sequence collecting the samples of said vertically folded impulse if said optional steps (Step 10.1.1 - Step 10.1.3) are carried out and of the sequence collecting the samples of said impulse response (X₀) referred to said Reference HULA in the considered delay bin if said optional steps are not carried out; said coefficients can be also computed by interpolating the elements of said spectrum (s_HULAifc); discarding one of the two solutions, which does not satisfy the inequalities: if b_H — 0 otherwise if said second order equation is solved; then the new estimate of T_H is obtained by adding said residual (A_H F_idft) to said estimate of T_H ; b) a second step wherein the new estimate of p is computed dividing said new estimate of T_H by F_idft ; c) a third step wherein said new estimate of T_H is employed to compute the new estimate of A as A = X_int('f_H) and X_int(T_H) is generated by interpolating the elements of said vector MS_HULAj0 where M is said oversampling factor employed in said IFFT calculations.

STDAE-S5) Overall folding and delay/amplitude estimation - In this step, the angular information (i.e., the spatial delays T_vi[l] and T_Hi[l]) computed in the previous steps are exploited to apply folding to the whole receive antenna array or to a portion of it. Similarly as the case of a MIMO FMCW radar, overall folding requires the computation of the iV₀-dimensional vectors X_O.OFM _/· X_I.OFM and X_2,OFM (see eq. (2.50)); the horizontal phase variation d as: where

If a peak is detected in the sequence of the absolute values of the elements of the vector X_O,OFM _A the CSDE algorithm is run to estimate, on the basis of the vectors X_O.OFM _/· XI.OFM and X2_,OFM_< the final estimates f_j [Z] and Ai[Z] of the parameters T_j[Z] and A_jfZ] characterizing the i-th target detected in the Z-th delay bin. Then, if the estimated delay f_j [Z] (see eq. (2.70)) is close to and the quantity ct_j [Z] appears in one of the couples of set S_b (2.23), it is discarded. Otherwise, the new couple (¾ [Z],F_{¾ ift}) (see eq. (2.55)) is added to the set S_b and the number of its elements (denoted by L_b) is increased by one.

In summary, the above method can be described as follows: said optional estimate of the vertical spatial delay (T_v) and said estimate of the horizontal spatial delay (T_H) are employed to compute the phase difference between said reference antenna and said virtual antennas of said optional sub-set of HULAs to be used in said constructive combination (Step 10.1.6 of Figure

29) to compute said overall folded impulse response (X_OF_,O) and its first three derivatives ({X_0Ffc;k = 1,2,3}) by combining the impulse reponse contributions referred to said sub-set of virtual antennas. Said checking (see Step 10.1.6 of Figure 29) is performed on the absolute value of said folded impulse response (X_0Fj0 ).

In summary, the method for the refinement (see Step 10.1.6 of Figure 29) can be described as follows (see Steps 10.1.6.1.0 - 10.1.6.1.2 of Figure 43):

- (Step 10.1.6.1.0) representation of a target normalized delay T = tAf, where Af is the frequency step size of the employed radar, as sum of a main portion (QCF_idft) and a residual portion (5FIDFT)_/· where F_IDFT = l/iV₀, where N₀ is the IFFT order, wherein said main portion is determined by searching the index (a) of the vector element X_0F o,a corresponding to the maximum absolute value of the elements of the vector X_OFO and said residual delay is found through an iterative procedure together with the corresponding target complex amplitude A = aexp(-jxp), where a is said target amplitude and y is said target phase;

(Step 10.1.6.1.1) Said iterative procedure includes a preliminary initialization comprising: setting the iteration index i to 1, the initial estimate of A equal to said vector element (X_OF,o_,a)_r the initial estimate of A = 2pdF_IDFT to zero and the initial estimate of T to said main portion (aF_IDFT); then iteratively

- (Step 10.1.6.1.2) the i-th iteration is fed by the estimates of D, T and A computed in the previous iteration, to generate new estimates of D, T and A, for a predetermined number of iterations, wherein each iteration includes: a) a first step where the residual D is calculated solving the first order equation Z¾D + c¾ = 0 OR the second order equation a_dD² + Z¾D + C_Q = 0 in the variable D with where ¾{x} denotes the real part and 3{x} denotes the imaginary part of the complex quantity x; discarding one of the two solutions, which does not satisfy the inequalities: ifa = 0 otherwise if said second order equation is solved; then the new estimate of T is obtained by adding said residual (A F_idft) to said main portion (af_IDFT); b) a second step wherein said new estimate of T is employed to compute the new estimate of A as A = Xi_nt(f) and Xi_nt(T) is generated by interpolating the elements of said vector MX_{OF O} where M is said oversampling factor employed in said IFFT calculations.

The above mentioned Steps 10.1.6.1.1 - 10.1.6.1.2 of Figure 43 can be replaced by an alternative method including (see Steps 10.1.6.2.1 - 10.1.6.2.2 of Figure 44):

(Step 10.1.6.2.1) Said iterative procedure includes a preliminary initialization including: a) setting the iteration index i to 1, the initial estimate of A equal to said vector element (X_OF_,O,¾)_r the initial estimate of

A = 2pdF_IDFT to zero and the initial estimate of T to said main portion (a F_IDFT) ; b) computing the real quantity p = F/F_IDFT; then iteratively - (Step 10.1.6.2.2) the i-th iteration is fed by the estimates of D, T and A and the real quantity p computed in the previous iteration, to generate new estimates of D, T, A and p, for a predetermined number of iterations, wherein each iteration includes a) a first step where the residual D is calculated solving the first order equation b_%A+ c¾ = 0 OR the second order equation where ¾{x} denotes the real part, 3{x} denotes the imaginary part of the complex quantity x [X_0Fkp;k = 1,2,3} is computed by interpolating the elements of said overall folded impulse response (Xo_F,fc)'‘ discarding one of the two solutions, which does not satisfy the inequalities: ifa = 0 otherwise if said second order equation is solved; then the new estimate of T is obtained by adding said residual (AF_idft) to said estimate of T; b) a second step wherein the new estimate of p is computed dividing said new estimate of T by F_idft; c) a third step wherein said new estimate of T is employed to compute the new estimate of A as A = Xi_nt{f) and Xi_nt(T) is generated by interpolating the elements of said vector MX_OFO where M is said oversampling factor employed in said IFFT calculations.

STDAEC-S2) Target cancellation - The contribution C^[Z], given by the i-th target detected in the Z-th time bin, to the vector X®[Z] is computed by means of eq. (4.223)(see Paragraph 4.1.6) and is cancelled from X®[Z], Cancellation consists in the evaluation of the new residual according to eq. (2.57).

In summary, the above method can be described as follows: said cancellation (see Step 10.2 of Figure 29) includes the computation and subtraction of the contribution given by said dominant point target to the impulse response contribution referred to said considered delay bin to generate a new residual of the impulse response and of the corresponding derivatives.

STDAEC-S3) Residual energy test - In this step, the energy £^'^⁺¹-⁾[Z] of the residual X^ⁱ⁺¹^[Z] evaluated in the previous step is computed on the basis of eq. (2.58)) and is compared with a positive threshold; if this energy is below this threshold, the STDAEC algorithm stops, otherwise the recursion index i is increased by one and a new iteration is started by going back to STDAEC-S1.

All the target information acquired from the a_j-th frequency bin are collected in the set

(2.81)

Ά ± {{Ai[l],T_t [l\,a_t [l\,f_vi [l],T_Hi [l]);i= 0,1. D[l]- l} with 1= 0, 1, L_{b —} 1 (it is assumed that D[l] targets have been detected in that bin). In T3-S2) the evaluation of spatial coordinates (i?;[Z],f_ί [Z],0_j [/]) is accomplished in the same way as in the FMCW case, i.e. on the basis of eqs. (2.60)-(2.62). Finally, the available information are merged to generate the overall set J_t (2.63), which, generally speaking, can be represented as a cloud of L points; note that J_t results of the union of the sets {J®} and that the set is still defined by eq. (2.100).

In summary, the method for the evaluation of the spatial coordinates can be summarized as follows: if the optional steps (Step 10.1.1 - Step 10.1.3 of Figure 29) are carried out, said computation (Step 11 of Figure 27) includes the computation of the range ( R) , the elevation (f ) and the azimuth (Q) as: and respectively, and, if the optional steps (Step 10.1.1 - Step 10.1.3 of Figure 29) are not carried out, said computation (see Step 11 of Figure 27) includes the computation of the range (R) and the azimuth (Q) as: and respectively, here, D/ is the frequency step size of the employed radar, c is the speed of light, l is the radar wavelength, d_Vv is the distance between two adjacent virtual antennas of said VULAs and d_VH is the distance between two adjacent virtual antennas of said HULAs. 2.3.2. Embodiment #2

The processing accomplished in the second embodiment of the proposed method and described by the block diagram illustrated in Figure 12 can be easily adapted to a MIMO SFCW radar system; most of the required changes are the same as those illustrated in the previous paragraph for the embodiment #1. In particular, the vector r_v (1.6) is replaced by the vector x_v (1.22), the FFT operation by the IFFT operation, frequency bins by delay bins in the STDREC and STDAEC algorithms (and, consequently, the spatial frequencies F_Vii\l\ and ¾ ;[/] by the normalised delays T_Vti[l] and T_H,i[i\' respectively) and the estimation of the complex amplitude Ai[p, l] observed on the p-th RVA by the estimation of the complex gain Ai[p, l] (see Paragraph 2.3.1. Embodiment #1). Another important change concerns the SBC procedure and, in particular, the evaluation of the triad C®_v) appearing in eq. (2.68). In fact, the computation of these vectors is based on eqs. (4.238)-(4.240) (see Paragraph 4.2.2), that are different from eqs. (4.232)-(4.234) employed in a MIMO FMCW radar system. Concerning 2D imaging, a similar adaptation can be derived also for a SFCW radar system for both embodiments. In summary, as in the FMCW radar system, for the second embodiment, said acquisition (see Step 8 of Figure 27) includes the acquisition of target energy and wherein said execution (see Step 10 of Figure 27) is repeated sequentially over the acquired list, ordered according to said target energy, and wherein said generation of the overall image (Step 11 of Figure 27) is carried out when all the delay bins have been analyzed. Said cancellation (see Step 10.2 of Figure 29) includes the computation and subtraction of the contribution given by said dominant point

— (v) target to said impulse response (X₀ ) and to the corresponding derivatives to generate a new residual of the impulse response and of the corresponding derivatives.

3. Some technical issues In this section, some technical problems investigated in the implementation of our method on real radar devices are illustrated and the technical solutions we have developed are described.

3.1. Unequal response of the receive antenna array In the derivation of the two embodiments described in Paragraph

2.2.1. Embodiment #1 and in Paragraph 2.2.2. Embodiment #2, both for 3D and 2D imaging, it is always assumed that the model of the signal received on the v-th RVA is expressed by eq. (1.11) (the same assumption is also made for the SFCW radar system; e.g., see eq. (1.23)). This model holds if the amplitudes {aj of the L overlapped oscillations do not change from antenna to antenna. Our experiments accomplished on commercial colocated radar devices have evidenced that: a) such amplitudes are not constant across the whole array; b) their differences are influenced by the azimuth and the elevation of each target. This may originate from the differences in the multiple receive chains employed in each MIMO device and the mismatches in the receive antenna patterns. This problem, that affects the data collected through both FMCW and SFCW radar systems, can be mitigated by enriching the physical array with a set of surrounding passive antennas; in this case, the array is artificially extended with new antennas along all its sides, so that the behavior of all its active antennas becomes more uniform.

In principle, this phenomenon can be accounted in our study for by including its effects in our received signal model. For instance, the real signal model (3.1) developed for an FMCW radar system can be modified as where a_n(q, f) is an attenuation factor depending on the angular coordinates of the Z-th target and v indicates the virtual element associated with the pair (p, q) of physical antennas. Consequently, the complex amplitude associated with the Z-th target detectable on the v-th RVA becomes (see eq. (1.18)) Ignoring the presence of the factor a·_n(bi, fi) in the algorithms developed in this document has the following implications: a) An error is introduced by the STDAEC algorithm in the cancellation procedure (see its task 3, step 1) of the embodiment #1. Note, in particular, that the estimate C_j [Z], referring to the Z-th target detected in the ct_j-th bin and generated by the STDAE technique in its step 5, is computed after the overall spatial folding; consequently, its absolute value represents a sort of spatial average computed over all the involved RVAs. Moreover, only the phase variations of this complex gain are accounted for in the computation of the contribution of this target to the matrix X®[Z] (see eqs. (4.230)-(4.231)). Similar comments apply to the STDAEC technique illustrated in Paragraph 2.4. Adaptation to 2D Imaging for the case of 2D imaging (see eq. (2.98)). b) An error in the cancellation procedure is introduced by the SBC algorithm in task 3 of the embodiment #2. In fact, in the computation of the vectors appearing in eq. (2.68), the variations in the absolute value of the complex gain Q[Z] are not accounted for (see eqs. (4.232)-(4.234)).

In principle, if the functions {a_n(q,f)} were known for all the RVAs, their effect could be compensated for after evaluating the estimates (^'bi,fΐ) of the angular coordinates of the i-th target; in fact, this result could be achieved by replacing the estimate Ci[l] of the complex gain Q[Z] with

Ci[v,Z]= C_j [Z]a_n(bi,ø.). (3.4) in the evaluation of the term C^®[Z] appearing in eqs. (4.230)- (4.231) (see STDAEC-S2 in Paragraph 4.1). Estimating the function a_n(q,f), however, is a time consuming task, since it requires a proper measurement setup and an anechoic chamber. For this reason, the alternative method we propose for compensating for the spatial amplitude distortion introduced by the radar array in the absence of any prior knowledge about the function a_n(q,f) is based on the exploitation of deep learning techniques (see "C. M. Bishop, Pattern Recognition and Machine Learning, ser. Information Science and Statistics. New York, NY: Springer, 2006.") in the SPE and on the adoption of a data-driven approach (see "0. Simeone, "A very brief introduction to machine learning with applications to communication systems," IEEE Transactions on Cognitive Communications and Networking, vol. 4, no. 4, pp. 648-664, 2018." and "N. Shlezinger, R. Fu, and Y. C. Eldar, "Deep soft interference cancellation for mimo detection," in ICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2020, pp. 8881-8885."). Our approach is motivated by the fact that: a) Deep learning techniques can be employed to approximate complicated functions, that do not lend themselves to a simple parametric representation and without requiring particular expertise in data pre-processing. b) The data-driven approach allows to train different models on the basis of data collected in a real scenario or synthetically generated, without prior knowledge about the parametric representation of the considered problem. The training procedure is important, because, through it, the involved network acquires the ability to generate a correct prediction on the basis of never seen data available at its input.

If the proposed method is adopted in the embodiment #1 and a MIMO FMCW radar system is considered, the block diagram illustrated in is replaced by that shown in Figure 15. As it can be easily inferred from comparing the first figure with the second one, the STDAEC block is replaced by a new block, that implements a new algorithm, called deep STDAEC (DSTDAEC). The initialization procedure of the DSTAEC is the same of the STDAEC and consists in: 1) Setting the iteration index i to 1. 2) Setting c⁽°⁾[Z]AX[Z], (3.5)

3) Selecting a reference antenna, identified by p = P_R;

4) Selecting a reference HULA, consisting of N_HULA adjacent and horizontally aligned RVAs; in the following, we assume, without any loss of generality, that the reference HULA includes the reference antenna and, consequently, is identified by p = p_j, p_j +

1, ..., p_F, with 0 £ p, £ p_R £ p_F £ N_EL - 1; note that

The DSTDAEC algorithm consists of three steps (its r-th step is denoted DSTDAEC-Sr in the following) and the processing accomplished within each of them is sketched below for the Z-th frequency bin (with 1 = 0, 1, ..., L_b — 1).

DSTDAEC-S1) Detection of a new target and estimation of its angular parameters - This step is identical to the first step of the STDAEC technique. Therefore, the N_VH X N_vv matrix X®[Z] (see eq. (2.31)) is processed to detect the strongest target (i.e., the Z-th target) contributing to the Z-th frequency bin and to estimate its parameters (qi[1], fi[1], fi[l],Ai[l]). Moreover, the processing accomplished within this step is based on the STDAE algorithm, whose detailed description is provided in Paragraph 2.2.1. Embodiment #1.

DSTDAEC-S2) Estimation of the distorted amplitudes - In this step, a deep neural network is employed to predict the absolute value² of Ci[v,l] (i.e., to predict a new estimate of \Ci[v,l\\; see eq. (3.4)) on the basis of the absolute value of the amplitude C₍ [Z] (estimated in STDAE-S5 ) . This network is designed to solve a regression problem through a supervised learning approach. Network training is based on the dataset

© = {(¾ ,¾); q = 1,2. N_t], (3.6) collecting N_t distinct points; here, and are N_VR-dimensional vectors representing the input of the network and its response, respectively, in the q-th trial. In network training, the following assumptions are made: a) the vectors A_q and A_q are known for any q; b) in the q-trial, the radar system observes a single point target and the target is detected in the frequency bin t>i. It is important to note that: a) The evaluation of the elements {C_j [v,Z]= \p,q,/]} of the vector in vector A_q (3.7) is based on eq. (2.25) and on the estimate Ci[l] computed in STDAE-S5 ; for this reason we have that

² The network is used for amplitude prediction, since we assume that the antenna array of a MIMO radar introduces an amplitude distortion only. b) If the vector A_q (3.8) is known, an estimate |C_j[Z]| of the absolute value of Q[Z] can be also computed as

The architecture proposed for the deep neural network and illustrated in Figure 16 has been inspired by that of feed- forward neural networks (see Chapter 5 of "C. M. Bishop, Pattern Recognition and Machine Learning, ser. Information Science and Statistics. New York, NY: Springer, 2006."). This architecture is characterized by input and output layers of the same size (N_vr) and K hidden layers. Each layer is composed by neurons and each neuron provides a non-linear combination of its inputs to the next layer. The amount of layers and neurons in the network are hyperparameters that they can be freely chosen, so that the weights and bias in each neuron can be optimized. Note, however, that the size of each layer decreases from the input layer to the output layer (see "A. Romero, N. Balias, S. E. Kahou, A. Chassang, C. Gatta, and Y. Bengio, "Fitnets: Hints for thin deep nets," 2014."). During training, the network has the ability to tune its parameters in order to learn a compressed representation of the input data, so generating an estimate of the amplitude distortion a_n(q,f).

DSTDAEC-S3) Target cancellation - In this step, the contribution C®[Z], given by the i-th target detected in the Z-th frequency bin, to the matrix X®[Z] (2.31) is computed. Eqs. (4.230)-(4.231) can be still exploited; however, eq. (4.231) is modified as - RR) K,I[*]]} where | \p,q,Z]|= [v,Z]|and \Ci[v,Z]|is predicted by the employed neural network. Then, cancellation is accomplished by computing the new residual vector X^ⁱ⁺¹^[Z] on the basis of eq. (2.57).

DSTDAEC-S4) Residual energy test - The energy £^'^⁺¹-⁾[Z] (2.58) of the residual spectrum vector X^ⁱ⁺¹^[Z] (2.57) (see STDAEC-S3 ) is compared with the positive threshold ^_STDAEC · If this energy is below the threshold, the DSTDAEC algorithm stops; otherwise, the recursion index Z is increased by one and a new iteration is started by going back to DSTDAEC-S1 . All the target information referring to the a_j-th frequency bin are collected in the set 7] (2.59); note that the quantity |C_j [Z]| appearing in the RHS of the definition (2.59) is replaced by

KJZ]! (3.10).

The method for amplitude compensation illustrated for the embodiment #1 can be easily adapted to the embodiment #2 (see Figure 12). In this case, the STDAEC technique is replaced by the DSTDAEC technique.Moreover, the SBC procedure is still based on eq. (2.68); however, in the computation of the vectors on the basis of eqs. (4.200)-(4.202), the dependence of the absolute value of the quantity = C_j [v,Z] on the antenna index v has to be taken into account. In summary, the above methods for a MIMO FMCW radar can be described as follows: said compensation of the amplitude distortion (see Step 10.1.7 of Figure 29) is carried out employing a deep neural network designed to solve a regression problem through a supervised learning approach wherein network training is based on a dataset collecting the amplitude of a number of said point targets; the compensation of the complex amplitude is carried out by multiplying said refined estimate of the complex amplitude (see Step 10.1.6 of Figure 29) by said estimated amplitude distortion for each of said virtual antennas.

If said compensation of the amplitude distortion (see Step 10.1.7 of Figure 29) is carried out, said computation of the contribution given by said dominant target (see Step 10.2 of Figure 29) takes into account the said amplitude distortion computed for each of said virtual antennas.

In summary, the above methods for a MIMO SFCW radar can be described as follows: said compensation of the amplitude distortion (see Step 10.1.7 of Figure 29) is carried out employing a deep neural network designed to solve a regression problem through a supervised learning approach wherein network training is based on a dataset collecting the amplitude of a number of said point targets; the compensation of the complex amplitude is carried out by multiplying said refined estimate of the complex amplitude (see Step 10.1.6 of Figure 29) by said estimated amplitude distortion for each of said virtual antennas. If said compensation of the amplitude distortion (see Step 10.1.7 of Figure 29) is carried out, said computation of the contribution given by said dominant target (see Step 10.2 of Figure 29) takes into account said amplitude distortion computed for each of said virtual antennas.

3.2 . Adaptation to different array shapes

In the description of the algorithms employed in the two embodiments of the proposed method, a URA has been always assumed until now; however, these algorithms can be easily adapted to different shapes of the receive array. For instance, in our experimental work, a colocated MIMO radar equipped with the physical array shown in Figure 13 has been used; the corresponding virtual receive array is shown in Figure 14. The first two processing tasks of our embodiments are carried out on an antenna-by-antenna basis; therefore, they do not depend on the array shape. For this reason, in the remaining part of this paragraph, we concentrate on the third task only. As far the initialisation procedure is concerned, the following consideration can be made:

1) The array structure illustrated in Figure 14 can be treated as an URA if its gaps are zero-padded.

2) The reference VULA has been selected in a way to maximize the number of non-zero vertically aligned RVAs and, consequently, the number of RVAs contributing to the estimation of the elevation angle, as illustrated in Figure 14.

3) The reference HULA has been selected in the middle of the antenna array; this mitigates the effects of the errors affecting the estimate of normalized vertical frequency in the vertical folding procedure; in fact, such errors may have a significant impact on the contributions of the HULAs that are farther from the reference HULA (see eq. (2.43)).

As far as the STDAE algorithm is concerned, the array structure illustrated in Figure 14 influences the vertical folding procedure since it involves VULAs of different sizes. More specifically, in the i-th iteration of the STDAEC algorithm vertical folding is accomplished by computing the N_HULA ^~ dimensional vector (see the definition (2.31) and STDAE-S3) where with p = Pi, Pi+ 1, P_F; here, expressed by eq. (2.43) and N_v\p] is the overall number of RVAs contained in the p-th VULA.

3.3. Antenna Coupling In our description of the SFE, the CSFE and the CSDE algorithms, it has been always assumed that the minimum frequency of the useful component contained in the observed data sequence can be arbitrarily small. This is not always true. For instance, in commercial colocated FMCW MIMO radar systems, a strong interference is observed in the lower portion of the spectrum computed on all the receive antennas. This is due to the electromagnetic coupling that originates from the small distance between adjacent transmit and receive antennas (see "S. Sun, A. P. Petropulu, and H. V. Poor, "MIMO Radar for Advanced Driver- Assistance Systems and Autonomous Driving: Advantages and Challenges," IEEE Signal Processing Magazine, vol. 37, no. 4, pp. 98-117, Jul. 2020."). Because of this interference, any target whose range is below a certain threshold cannot be detected in a reliable fashion. This phenomenon is called mutual coupling (see "C. M. Schmid, S. Schuster, R. Feger, and A. Stelzer, "On the Effects of Calibration Errors and Mutual Coupling on the Beam Pattern of an Antenna Array, " IEEE Transactions on Antennas and Propagation, vol. 61, no. 8, pp. 4063-4072, Aug. 2013.").As far as we know, various methods based on calibration measurements have been proposed to solve this problem (see "C.M. Schmid, C. Pfeffer, R. Feger, and A. Stelzer, "An FMCW MIMO radar calibration and mutual coupling compensation approach," in 2013 European Radar Conference, Oct. 2013, pp. 13- 16.").

4. Detailed Description of the Algorithms Employed in the Proposed Embodiments

In this section, various mathematical details about the algorithms employed in the two embodiments illustrated in the previous section are provided. We first focus on the algorithms introduced in our description of the first embodiment. Then, we briefly show some mathematical results exploited in the second embodiment only.

4.1. Embodiment #1

Since a full description of the processing accomplished in task #1 (Tl) of the first embodiment has been provided in the previous section, in this paragraph we comment on the RVA selection accomplished in task #2 (T2) and provide a detailed description of the SFE (T2-S2), of the CSFE (T2-S2 and T3-S1), of the CSDE (T2-S2 and T2-S3) and of the target cancellation procedures employed in task #2, step 2 (T2-S3) and in task #3, step 1 (T3-

Sl).

4.1.1. RVA selection

The first step of T2 aims at selecting N_A distinct RVAs within the set of the N_VR available RVAs. The algorithm adopted in our computer simulations consists in randomly extracting N_A distinct integers from the set {0,1,..., N_VR 1}; these integers identify the selected antennas. It is important to note that, generally speaking, the exploitation of a subset of the available antennas is motivated by the need of reducing the computational effort required by T2 as much as possible. Moreover, even if a deterministic selection of N_A antennas is possible, this should be avoided since, when multiple consecutive snapshots are processed to generate indipendent images, the radar system can benefit from antenna diversity by changing the subset of N_A antennas from snapshot to snapshot in a random fashion. Of course, this algorithm is not employed if N_A = N_VR .

4.1.2. Single frequency estimator

The algorithm executed by the SFE processes the real time-domain vectors

(4.1)

T

^X0 — [^X0,0’^X0,1>^{■ ■ ■}>¾ N-l] (4.2) and

_T (4.3)

^X2 = [^c2,0>^c2,1>···>^c2,N-ΐ\ > where {x₀,n) is the sequence of available noisy measurements,

(4.4)

^L Y TTl m,n — ^{,l L}0,n> with m = 1, 2 and n = 0,1,.. N — 1 . This algorithm is derived under the assumption that the n-th element of the vector x₀ (4.1) can be expressed as

(4.5) xo,n = acos(2nf_x nT_s + y) + w_n, with n = 0,1,..., N — 1 . Based on the vectors x₀ (4.1), x_x (4.2) and X2 (4.3), the SFE computes an estimate / of the frequency f_x and an estimate C of the complex amplitude

(4.6) characterizing the sinusoid contained in the sequence {Xo,n}· The algorithm executed by the SFE operates as follows. The vectors x₀ (4.1), x_x (4.2) and x₂ (4.3) are zero padded by appending (M — l)iV zeros to each of them, where M is a integer parameter; this results in the MN-dimensional vectors (4.7) and

(4.9)

These vectors undergo DFT processing of order

(4.10)

N_{0 A} MN; this produces the N₀-dimensional vectors and respectively. Then, the vectors X₀ (4.11), X_x (4.12) and X₂ (4.13) are processed to compute the estimates / and C . It is important to point out that, on the one hand, the algorithm employed for estimating f_x relies on the observation that this frequency can be always expressed as where a_x is the index of the frequency bin and d_c is a residual such that

For this reason, estimating the frequency f_x is equivalent to computing the estimates a_x and d_c of a_x and d_c, respectively. Note also that the representation (4.14) can be also rewritten as where F_x = f_xT_s and with (see eq. (4.16)) ifa_x = 0

(4.19) ifa_x = iV_no— 1 otherwise or, equivalently,

_< -Ji^{<NAx£0 if}“x^{= iV}o-l· p 7G

— — < i\M_r < — otherwise v M M

Consequently, the range which the parameter D_c (i.e., d_c) belongs to can be made arbitrarily small by selecting iV₀ (that is, M) large enough (say, M 3;10). On the other hand, the algorithm devised for the estimation of the parameter C (4.6) relies on the fact that: a) in the absence of noise,

(4.21) where is the Fourier transform of the sequence {Xo_,n}/ b) since (see eq. (2.9)) (423) or, equivalently,

*o(/_fc) = MX _0,fc where

(4.25) for k = 0, 1, ..., iV₀ — 1, the vector

(4.26)

X_S=MX o, collects N₀ uniformly spaced samples of the function X₀(/) (4.22). For this reason, an approximate value of X₀(/*.) at a frequency f_x different from a multiple of 1/(iV₀T_s) can be computed by interpolating the elements of the vector X_s (4.27). In our computer simulations, the barycentric interpolation described in "J. Selva, "Efficient Wideband DOA Estimation Through Function Evaluation Techniques," IEEE Trans. Sig. Proc., vol. 66, no. 12, pp. 3112-3123, June 2018." has been used; however, the use of other interpolation schemes is also possible.

The SFE computes the estimates a_x, d_c and C as follows. First of all, the estimate a_x is evaluated by means of the periodogram method (e.g., see "J. G. Proakis and D. G. Manolakis, Digital Signal processing: Principles, Algorithms and Applications. Third Edition. New York, Ny: Prentice Hall, New York, 1993."), i.e. as

(4.28) ry = are max ae{0,l.N₀/2] K 4

Then, an iterative algorithm is employed to generate the estimates d_c and C . The i-th iteration is fed by the estimates of A_x, f_x and C, respectively, and generates the new estimates The algorithm is initialised by setting the iteration index i to 1,

(4.29) and by computing the complex coefficients ^(2^), K₂(2 a_x) and K₃(2 a_x) where

(4.30) In its i-th iteration (with i = 1,2,.. iV_SFE, where iV_SFE is the overall number of iterations), it accomplishes the two steps described below .

1) The equation

(4.31) or the equation (4.32) in the variable D_c are solved; here,

(433)

If eq. (4.31) is solved, one of its two solutions can be easily discarded, since it does not satisfy the inequalities (4.19). The solution of eq. (4.32) (or the selected solution of eq. (4.31)) represents D_c\ for this, reason, it is used to compute the new estimate (see eq. (4.14) and the definition (4.18))

(4.36)

Of f_x-

2 The estimate / (4.36) is employed to compute the new estimate where is generated by interpolating the elements of the vector X_s (4.26).

It important to point out that: a) The coefficients /^(x), K₂(x) and K₃(x) employed in eqs. (4.33)- (4.35) can be computed for any x 6 {0 , 1, ..., N₀ — 1} and stored in a memory. b) Eq. (4.32) is simpler to solve than eq. (4.31), but is expected to provide a less accurate estimate of D_c . If eq. (4.31) is used, the solution is selected. In our simulation results illustrated in Section 6, eq. (4.32) has been employed. c) The SFE can be exploited even if the sinusoid appearing in the RHS of eq. (4.105) is replaced by a sum of L sinusoids, characterized by distinct frequencies (see eq. (1.12)). In this case, one of the L sinusoids is considered as the useful component of the processed sequence and the remaining ones as part of the noise affecting it. Note that, in this case, the SFE estimates the parameters of the sinusoid generating the highest peak in the amplitude spectrum of the noisy sequence {x₀,n} and the quality of its estimates is affected by both the amplitudes and the frequencies of the other (L— 1) sinusoids. d) The SFE algorithm can be easily extended in a way that multiple frequencies are detected and estimated in parallel; the approach to be followed is the same as that illustrated in Sec. IV of "J. Selva, "ML Estimation and Detection of Multiple Frequencies Through Periodogram Estimate Refinement," IEEE Sig. Proc. Lett., vol. 24, no. 3, pp. 249-253, 2017.". Note that, in this case, the frequencies are searched for by setting a proper threshold and verifying that the frequency spacing between frequencies is large enough (so that appreciable mutual interference is avoided). Moreover, a proper threshold on the amplitude | (i.e., the radar cross section of the target) is set to estimate the parameters of targets with comparable RCS.

An alternative to the SFE is represented by a related estimation algorithm (called single frequency estimator #2, SFE#2) that relies on the same mathematical results as the SFE and has the same structure. However, it is also based on the idea of taking the estimate F_x ^generated in its (i— l)-th iteration as a coarse estimate of F in the following (i.e., the i-th) iteration. The initialization phase of SFE#2 is very similar to that of SFE, the only difference being represented by the fact that the quantity

(4.39) is also defined. The new estimate D_c of D_c is computed through eq. (4.31) or eq. (4.32); in the evaluation of the coefficients ^C _S^} ^or {^4^’^cs^} appearing in the RHS of these equations, the quantities [K_p(2a_x);p = 1,2,3] are replaced by and respectively, for p =1,2,3 and k = 1,2,3, where

(4.42)

Then, the new estimate of the normalised frequency s computed as The new estimate of C is computed in the same way as in the SFE. It is important to point out that the comments b) to d) also apply to SFE#2 and that: a) The quantities required in the first step of SFE#2 can be computed exactly through eqs. (4.40) and (4.41), respectively. However, they can be also evaluated, in an approximate fashion, at a lower computational cost by interpolating I consecutive elements of the iV₀-dimensional vector

(4.44)

K_p = [K_p(0),K_p(l). K_p(iV₀ — i)]^T . and of the N₀-dimensional vector X₀ (4.11), X_x (4.12) and X₂ (4.13), respectively, where I denotes the adopted interpolation order. b) the quantity (4.42) accounts for both the components of F_x ^ (i.e., for both the integer a_x and the residual D_c^ ) appearing in the RHS of (4.36).

In the remaining part of this paragraph, the equations on which the SFE and SFE#2 previously described are based are illustrated. We consider the signal model expressed by eq. (4.5), i.e. x_n = acos(2nnf T_s + ip) + w_n (4.45) and rewrite it as x_n = Cexp(j2nnF)+ C^*exp(—j2nnF) + w_n, (4.46) where and

F A fT_s. (4.48)

Our objective is estimating the parameters C and F of the sinusoid contained in the sequence { x_n; n = 0,1,..., N — 1); if the sequence {w_n} can be modelled as white Gaussian noise, optimal estimation of these parameters in the maximum likelihood (ML) sense is obtained by minimizing the mean square error (MSE) with respect to the trial parameters F and C; here,

E_n(F, C) ± [_Xn - s_n(F, C)]² (4.50) is the square error between the noisy sample x_n and the sample s_n(F,i4)= Cexp that represents the useful component of x_n (4.46) under the assumption that F = F and C = C . Substituting the RHS of the last equation in the RHS of eq. (4.50) yields where C_R = ¾{C}, C_j = 3{C} and f T _n = 2nnF. (4.53)

Then, substituting eq. (4.52) in the RHS of (4.49) produces, after some manipulation, where X_R(F)A ¾{X(F)}, X,(F) A 3{X(F)}_r

9_R(J)= ¾M F)}, 5_/(F)= %{g(F)} and

The maximum likelihood (ML) estimates F_ML and C_ML of the F and C, respectively, can be computed by solving the optimization problem

{FML-FML)= argmijiF(F,C). (4.58)

Solving this problem is not easy because of the nonlinear dependence of the function F(/,C) (4.54) on F. For this reason, an iterative method, known as alternating minimization (AM) is adopted in our work to solve it (e.g., see "I. Ziskind and M. Wax, "Maximum likelihood localization of multiple sources by alternating projection," IEEE Trans. Ac., Speech and Sig. Proc., vol. 36, no. 10, pp. 1553-1560, Oct. 1988."). In practice, this method is initialised by computing a coarse estimate of C

~(o) ~(o) and a coarse estimate F of F under the assumption that C = C

Then, an iterative procedure for progressively refining these estimates is executed. At the beginning of the i-th iteration (with i= l,2,...), the estimate of F evaluated in the previous iteration is available and, based on this, the following new

~(i) estimates are computed: a) a new estimate (denoted C ) of C conditioned on F = F^ \ b) a new estimate (denoted F^ ) of F conditioned on . Developing an algorithm based on this approach requires solving the following two problems: PI) the minimization of the function F(F,C) with respect to C for a given F; P2) the minimization of F(F,C) with respect to F for a given C. Let us focus first on problem PI. The minimum of the function F(F,C) (4.54) with respect to C is the solution of the equations

C_R -X_R(F)+ C_R g_R(F)+ C,g,(F)= 0 (4.59) and

C,-X,(F)- C,g_R(F)+ C_R g,(F)= 0, (4.60) that result from computing the partial derivative of F(F,C) with respect to C_R and C_Ir respectively. These equations can be easily rewritten as

[1+ 3_R(F)]C_R + g,(F)C,= X_R(F) (4.61) g,(F)C_R + [l-g_R(F)]C,=X,(F) (4.62) Solving this linear system of equations in the unknowns C_R and C_j produces

Therefore, for a given F, the function F(F,C) (4.54) is minimized with respect to C by setting C = C, where C is computed on the basis of eq. (4.63). Note that the last formulas are exact and require the evaluation of the two functions X(F) (4.56) and g(F) (4.57). On the one hand, an efficient computation of the function g(F) can be accomplished by resorting to the simple and exact formula with w = exp(—]4nF). On the other hand, the procedure we adopt for the evaluation of X(F) is approximate, being based on the use of the DFT and of the interpolation technique as described in Paragraph 4.1.2.

Then, let us focus on problem P2 . Unluckily, no exact solution is available for this problem. The approximate solution we propose is based on: 1) assuming that the variable F can be expressed as (see eq. (4.17)) a D

F = (4.65) T₀ ⁺ 2p’ where a is a known integer and D is an unknown quantity satisfying the inequalities (see eq. (4.19)) if a = 0 if a = iV₀ — 1. (4.66) otherwise

2) approximating the trigonometric functions appearing in the RHS of eq. (4.52) (with k = 1 and 2 ) . More specifically, based on eq. (4.65), the function E_n(F,C) (4.52) is rewritten as

Then, eq. (4.67) can be put in the equivalent form by exploiting standard trigonometric formulas. Finally, substituting the RHS of eq. (4.69) in that of eq. (4.49) yields

The last formula depends on the variable D through the nonlinear functions sin(nqD) and cos^q^), with q = 1 and 2. However, based on the inequalities (4.66) and on the assumption that the parameter M is large enough (say, M 3;10)), the following standard approximations and all based on the Taylor series of the trigonometric functions sin(x) and cos(x), can be adopted. Then, substituting the RHS of eqs. (4.71)-(4.74) in that of eq. (4.70) yields, after some manipulation and K_v(x) are defined by eq. (4.30). Note that the quantities X_{Q Q}, X^_Q and X_2Q represent the a-th element of the vectors X₀ (4.11), X_x (4.12) and X₂ (4.13), respectively. Given the function £SFE_,CI (4-75), an approximate expression for D is obtained by computing the partial derivative of this function with respect to D and setting it to zero; this results in The last equation can be easily put in the form which is equivalent to aD² + bD + c= 0, (4.81) where and

It should be expected that only one of the two solutions of Eq. (4.81) satisfies the inequalities (4.66); this solution is denoted A and is used to compute the estimate (see eq. (4.65)) a A

F = (4.85) T₀ ⁺ 2p of F. This solves the second problem.

4.1.3. Complex single frequency estimator

The algorithm executed by the CFSE processes the complex vectors (4.86) with m 0,1,2 and 3; here, {r_0,n} is the sequence of available noisy measurements,

(4.87)

^A g m,n ^{— ,L} _iTΪC _m L0,n> with m = 1,2 and 3 and n = 0,1,.. N 1 This algorithm is derived under the assumption that the n-th element of the vector x₀ (4.86) can be expressed as

(4.88) r_0,n = aexp(j(2nF_xn + f_c))+ w_n> with n = 0,1,...,N — 1; here, F_x is a normalized frequency, a is a real positive parameter and {w_n} is a complex noise sequence. It is worth stressing that F_x represents a normalized frequency if this algorithm is applied to the time domain complex samples acquired by an FMCW radar and a normalized spatial frequency if it is applied to the FFT coefficients computed over multiple receive antennas of the same radar device for the estimation of the angular coordinates of targets.

The CSFE algorithm computes an estimate F_x of the frequency F_x and an estimate A_x of the complex amplitude

(4.89)

A_{x A} aexp(j<P_x) characterizing the complex exponential contained in the sequence

{-*^•0,n}·

It operates as follows. The vectors {x₀, x_lr x₂, X₃} (4.86) are zero padded by appending to it (M — l)iV zeros, where M is an integer parameter (oversampling factor); this results in the MN- dimensional vectors

X -fix FoL f (4.90) xm,ZP [l^xml ^U(M-1)NJ ’ with m = 0,1,2 and 3. These vectors undergo N₀=MN-th order DFT processing; this produces the iV₀-dimensional vector with m = 0,1,2 and 3 . Note that, generally speaking, the frequency

F_x is not a multiple of the fundamental frequency l/iV₀ associated with the FFT processing; for this reason, it can be expressed as where a_x is an integer belonging to the set {0 , 1, ..., iV₀ — 1} and ¾ is a real quantity, such that < 0.5 for a_x = 0 (4.93) d_c < 0 for a_x = N₀ — 1. d_c < 0.5 otherwise

The representation (4.92) can be also rewritten as

2p (4.94)

2nF_x — _ oc_x + D_c,

N o where with (see eq. (4.93)) for a_x = 0 — for a_x = N₀ — 1 (4.96) otherwise The CSFE computes the estimates and of the parameters A_x, a_x and D_c, respectively; this allows to compute the estimate of the normalized frequency F_x (4.92).

The CSFE operates as follows. First of all, the estimate a_x is evaluated by means of the periodogram method (e.g., see "J. G. Proakis and D. G. Manolakis, Digital Signal processing: Principles, Algorithms and Applications. Third Edition. New York, Ny: Prentice Hall, New York, 1993."), i.e. as

Then, an iterative algorithm is employed to generate the estimates . The i-th iteration is fed by the estimates ) and , respectively, and generates ~(i) _^(i) -(i) the new estimates A_x , F_x and A_x . The algorithm is initialised by setting the iteration index i to 1,

In its i-th iteration (with i= 1,2,...,iV_CSFE, where iV_CSFE is the overall number of recursions), the two steps described below are accomplished.

1) The equation (4.100) or the equation

(4.101) in the variable are solved; here (see eqs. (4.139)- (4.141)),

If eq. (4.100) is solved, one of its two solutions can be easily discarded, since it does not satisfy the inequalities (4.93). The solution of eq. (4.101) (or the selected solution of eq.

(4.100)) represents for this, reason, it is exploited to compute the new estimate (see eq. (4.97))

(4.105) of F_x.

-(0 2) The estimate F_x is employed to compute the new estimate and is generated by interpolating the elements of the vector that collects N₀ uniformly spaced samples of the spectrum of the vector x₀ (4.86); in our work, similarly as the SFE, barycentric interpolation is used.

It important to point out that: a) Eq. (4.101) is simpler to solve than eq. (4.100), but is expected to provide a less accurate estimate of D_c . If eq. (4.100) is used, the solution is selected. In our simulation results illustrated in Section 6, eq. (4.101) is used.

b) The CSFE can be exploited even if the exponential appearing in the RHS of eq. (4.88) is replaced by a sum of L exponentials, characterized by distinct frequencies (see eq. (1.23)). This is equivalent to considering one of the L exponentials as the useful component of the processed signal and the remaining ones as part of the noise affecting it. Note that, in this case, the CSFE estimates the parameters of the complex exponential generating the highest peak in the amplitude spectrum of the noisy sequence { x_0n} and the quality of its estimates is affected by both the amplitudes and the frequencies of the other (L— 1) exponentials. An alternative to the CSFE is represented by a related estimation algorithm (called complex single frequency estimator #2, CSFE#2) that relies on the same mathematical results as CSFE and has the same structure. However, it is also based on the idea of taking the estimate F_x ^generated in its (i— l)-th iteration as a coarse estimate of F in the following (i.e., the i-th) iteration. The initialization phase of CSFE#2 is very similar to that of CSFE, the only difference being represented by the fact that the quantity

(4.109) is also defined. The new estimate D_c of D_c is computed through eq. (4.100) or eq. (4.101); in the evaluation of the coefficients ^C _S^} ^or {^4^’^cs^} appearing in the RHS of these equations, the quantity {¾ ¾.; k = 1,2,3] is replaced by for k = 1,2,3, where

(4.111)

Then, the new estimate of the normalised frequency is computed as

(4.112)

The new estimate of C is computed in the same way as in the CSFE. It is important to point out that all the comments illustrated for the CSFE also apply to CSFE#2 and that: a) The quantity k = 1,2,3j required in the first step of

CSFE#2 can be computed exactly through eq. (4.110). However, it can be also evaluated, in an approximate fashion, at a lower computational cost by interpolating I consecutive elements of the iV₀-dimensional vectors (X_m; m = 1,2,3} (4.91), where I denotes the adopted interpolation order. b) the quantity (4.111) accounts for both the components of

F_x ^ (i.e., for both the integer a_x and the residual ) appearing in the RHS of (4.105).

In the remaining part of this paragraph, the equations on which the CSFE and CSFE#2 previously described are based are illustrated. We consider the complex counterpart of the real model expressed by eq. (4.45), i.e. x_n = aexp(j (2nnfT_s + ø))+ w_n (4.113) and rewrite it as x_n = Aexp(j2nnF)+ w_n, (4.114) where

A = aexpQ^'cfi) (4.115) and

F ^A fT_s. (4.116)

Our objective is estimating the parameters A and F of the complex exponential contained in the sequence {x_n; n = 0,1,...,N — 1); if the sequence {w_n} can be modelled as white Gaussian noise, optimal estimation of these parameters in the maximum likelihood (ML) sense is obtained by minimizing the mean square error (MSE) with respect to the trial parameters F and A; here is the square error between the noisy sample x_n and the sample that represents the useful component of x_n (4.114) under the assumption that F = F and A = A. Substituting the RHS of the last equation in that of eq. (4.118) yields f_h = 2phY. (4.121) Then, substituting the RHS of eq. (4.120) in that of eq. (4.117) produces, after some manipulation,

The procedure we follow to minimise the function E(F,A) (4.122) with respect to F and A is similar to that illustrated in Paragraph 4.1.2 for the SFE. For this reason, we first show how the function E(E,A) (4.122) can be minimised with respect to A for a given F; then, we show how the same function can be minimised with respect to F for a given A. The solution of the first problem can be obtained by solving the equations

AR— Xo_,a,R = 0 (4.125) and

A,-Co_,a,,=0, (4.126) that result from computing the partial derivative of the function E(E,A) (4.122) with respect to A_R and A_j, respectively. Solving this linear system of equations in the unknowns A_R and A_j produces

A=X₀fi. (4.127)

Note that the last formula is exact and require the evaluation of the function X_0Q (see the definition (4.124)).

Unluckily, an exact expression for the value of F minimising the function E(E,A) for a given A cannot be derived because of its nonlinear dependence on this variable. The approximate solution we propose is based on: 1) assuming that (see eqs. (4.94)-(4.96)) where N₀ =MN, a is a known quantity, whereas A is a variable satisfying the inequalities for a = 0 — for a = ^jV₀ - 1 (4.129) otherwise

2) approximating the trigonometric functions appearing in the RHS of eq. (4.120). More specifically, based on eq. (4.128), the function E_n(F,A) (4.120) is rewritten as where Then, eq. (4.130) is put in the equivalent form

E_n(F,A) = O_n I²+A_R ²+A_I ²

-20 _n,R ¾ + ^x _n,i 2i_/)[cos(0_n)cos(n2) - sin(0_n)sin(n 2)] (4.132)

-20_n,/ ¾ - *_n,/? ¾)[sin(0_n)cos(n2) + cos(0_n)sin(n2)] by exploiting standard trigonometric formulas. Then, substituting the RHS of eq. (4.132) in that of eq. (4.117) yields E{F,A) — E_x + A_R + A_j

N-l

-2 — V ( A;,,„ ¾ + ¾, ?[, )[cos(e„)cos(n2) - sln(e„)sln(ti2 (4.133 -2l5¹(_¾2_B-¾„2,)[_Si„(e„)cos(„2) ₊ co_S(9„)s.„(n2⁾ n-0

The last formula depends on D through the nonlinear functions sin(nil) and cos(nil) . However, based on the inequalities (4.129) and on the assumption that M is large enough (see eq. (2.37)), the following standard approximations sin(hD)—hD—h³ — (4.134) and all based on the Taylor series of the trigonometric functions sin(x) and cos(x), can be adopted.

Finally, substituting the RHSs of eqs. (4.134) and (4.135) in that of eq. (4.133) yields, after some manipulation The function £cs_FE,a(^ ,A^') (4.136) can be minimized with respect to A by taking its partial derivative with respect to this variable and setting it to zero; this results in the equation that can be easily put in form aA² + bA + c= 0 (4.138) where and

(4.141)

It should be expected that only one of the two solutions of Eq. (4.138) satisfies the inequalities (4.129)³ . If this solution is denoted A , the estimate (see eq. (4.128))

³ In running our computer simulations, only the first-order problem has been solved. a D

F = (4.142) N₀ 2p of F can be computed.

4.1.4. Complex single delay estimator

The algorithm executed by the CSDE processes the complex vectors with m = 0,l,2 and 3; here, {r_0,n} is the sequence of available noisy measurements,

(4.144)

^{L l}TTl m,n ^{~ ,} _m L0,n> with 777= 1,2 and 3 and 77= 0,1,...,N 1 This algorithm is derived under the assumption that the 77-th element of the vector x₀ (4.143) can be expressed as

(4.145) with 7i= 0,1,...,N 1; here, T_x is a normalized delay, a is a real positive parameter and {w_n} is a complex noise sequence. It is worth stressing that T_x represents a normalized delay if this algorithm is applied to the frequency response samples acquired by an SFCW radar and a normalized spatial delay if it is applied to the IFFT coefficients computed over multiple receive antennas of the same radar device for the estimation of the angular coordinates of targets.

The CSDE algorithm computes an estimate T_x of the delay T_x and an estimate A_x of the complex amplitude

(4.146)

A_{x A} aexp(—/#>*) characterizing the complex exponential contained in the sequence

{-^0,n}·

It operates as follows. The vectors {x₀, x_lr x₂, X3} (4.143) are zero padded by appending to it (M — l)iV zeros, where M is an integer parameter (oversampling factor); this results in the MN- dimensional vector with m = 0,1,2 and 3. The four vectors {x_m,zp} undergo N₀=MN-th order IDFT processing; this produces the iV₀-dimensional vector

(4.148) with m = 0,1,2 and 3. Note that, generally speaking, the delay T_x is not a multiple of the fundamental delay l/iV₀ associated with the IFFT processing; for this reason, it can be expressed as

(4.149) where a_x is an integer belonging to the set {0, 1, No— 1} and d_c is a real quantity, such that 0.5 for a_x = 0 (4.150) c < 0 for a_x = N₀ — 1. _c < 0.5 otherwise

The representation (4.149) can be also rewritten as where with (see eq. (4.150)) for a_x = 0 for a_x = N₀ — 1 (4.153) otherwise

The CSDE computes the estimates A_x, a_x and D_c of the parameters A_x, a_x and D_c, respectively; this allows to compute the estimate

(4.154) of the normalized delay T_x (4.149). The CSDE operates as follows. First of all, the estimate a_x is evaluated by means of the periodogram method (e.g., see "J. G. Proakis and D. G. Manolakis, Digital Signal processing: Principles, Algorithms and Applications. Third Edition. New York, Ny: Prentice Hall, New York, 1993."), i.e. as

(4.155)

Then, an iterative algorithm is employed to generate the estimates D_c and A_x . The i-th iteration is fed by the estimates ~(i-i) ~(i-i) -(i- 1)

D_c , T_x and A_x of D_c, T_x and A_x, respectively, and generates

~(i) ~(i) ~(i) the new estimates D_c , T_x and A_x . The algorithm is initialised by setting the iteration index i to 1 and

In its i-th iteration (with i= 1,2,...,iV_CSDE, where iV_CSDE is the overall number of recursions), the two steps described below are accomplished.

1) The equation

(4.157) or the equation

(4.158) in the variable D_c are solved; here (see eqs. (4.196)-(4.198)), If eq. (4.157) is solved, one of its two solutions can be easily discarded, since it does not satisfy the inequalities (4.150). The solution of eq. (4.158) (or the selected solution of eq.

~(i)

(4.157)) represents D_c ; for this, reason, it is exploited to compute the new estimate (see eq. (4.154))

(4.162) of T_x.

~(i) ~(i)

2) The estimate T_x is employed to compute the new estimate A_x =

A_X,R+jA_XJ as

(4.163) and generated by interpolating the elements of the vector

X_s = M X₀, (4.164) that collects N₀ uniformly spaced samples of the impulse response of the vector x₀ (4.143); in our work, similarly as the CSFE, barycentric interpolation is used.

It important to point out that: a) Eq. (4.158) is simpler to solve than eq. (4.157), but is expected to provide a less accurate estimate of D_c . If eq. (4.157) is used, the solution is selected. In our simulation results illustrated in Section 6, eq. (4.158) is used. b) The CSDE can be exploited even if the exponential appearing in the RHS of eq. (4.145) is replaced by a sum of L exponentials, characterized by distinct delays (see eq. (1.23)). This is equivalent to considering one of the L exponentials as the useful component of the processed signal and the remaining ones as part of the noise affecting it. Note that, in this case, the CSDE estimates the parameters of the complex exponential generating the highest peak in the amplitude of the impulse response produced by the IDFT of the noisy sequence { x_0n} and the quality of its estimates is affected by both the amplitudes and the delays characterizing the other (L— 1) exponentials.

An alternative to the CSDE is represented by a related estimation algorithm (called complex single delay estimator #2, CSDE#2) that relies on the same mathematical results as CSDE and has the same structure. However, it is also based on the idea of taking the estimate T_x ^generated in its (i— l)-th iteration as a coarse estimate of T in the following (i.e., the i-th) iteration. The initialization phase of CSDE#2 is very similar to that of CSDE, the only difference being represented by the fact that the quantity

(4.166) is also defined. The new estimate D_c of D_c is computed through eq. (4.157) or eq. (4.158); in the evaluation of the coefficients C_Q ⁾j appearing in the RHS of these equations, the quantity {¾ ¾.;k = 1,2,3] is replaced by for k = 1,2,3, where

(4.168) Then, the new estimate of the normalised delay T_x is computed as

(4.169)

~(£)

The new estimate A of A is computed in the same way as in the CSDE. It is important to point out that all the comments illustrated for CSDE also apply to CSDE#2 and that: a) The quantity k = 1,2,3j required in the first step of CSDE#2 can be computed exactly through eq. (4.167). However, it can be also evaluated, in an approximate fashion, at a lower computational cost by interpolating / consecutive elements of the iV₀-dimensional vectors (X_m; m = 1,2,3} (4.148), where I denotes the adopted interpolation order. b) the quantity (4.168) accounts for both the components of

T_c ^ (i.e., for both the integer a_x and the residual A_x^) appearing in the RHS of (4.162).

In the remaining part of this paragraph, the equations on which the CSDE and CSDE#2 previously described are based are illustrated. We consider the complex counterpart of the real model expressed by eq. (4.45), i.e. x_n = aexp(—j(2nnTT_s + ø))+ w_n (4.170) and rewrite it as x_n = Aexp(—j2nnT) + w_n, (4.171) where

A = aexp(—/ø) (4.172) and

T = x Af. (4.173)

Our objective is estimating the parameters A and T of the complex exponential contained in the sequence {x_n; n = 0,1,...,N — 1); if the sequence {w_n} can be modelled as white Gaussian noise, optimal estimation of these parameters in the maximum likelihood (ML) sense is obtained by minimizing the mean square error (MSE) with respect to the trial parameters T and A; here is the square error between the noisy sample x_n and the sample that represents the useful component of x_n (4.170) under the assumption that T = T and A = A. Substituting the RHS of the last equation in that of eq. (4.175) yields f = 2phT. (4.178)

Then, substituting the RHS of eq. (4.177) in that of eq. (4.174) produces, after some manipulation,

The procedure we follow to minimise the function E(T,A) (4.179) with respect to T and A is similar to that illustrated in Paragraph 4.1.2 for the SFE. For this reason, we first show how the function E(T,A) (4.179) can be minimised with respect to A for a given T; then, we show how the same function can be minimised with respect to T for a given A. The solution of the first problem is computed by solving the equations and that result from computing the partial derivative of the function E(E,A) (4.179) with respect to A_R and A_], respectively. Solving this linear system of equations in the unknowns A_R and A_j produces (4.184)

Note that the last formula is exact and require the evaluation of the function X₀a (see the definition (4.181)).

Unluckily, an exact expression for the value of minimising the function for a given cannot be derived because of its nonlinear dependence on this variable. The approximate solution we propose is based on: 1) assuming that (see eqs. (4.151)

(4.153)) (4.185) where is a known quantity, whereas is a variable satisfying the inequalities (4.186) 2) approximating the trigonometric functions appearing in the RHS of eq. (4.177). More specifically, based on eq. (4.185), the function (4.177) is rewritten as where (4.188)

Then, eq. (4.187) is put in the equivalent form by exploiting standard trigonometric formulas. Then, substituting the RHS of eq. (4.189) in that of eq. (4.174) yields

The last formula depends on D through the nonlinear functions and However, based on the inequalities (4.186) and on the assumption that oversampling factor M is large enough, the following standard approximations and all based on the Taylor series of the trigonometric functions sin(x) and cos(x), can be adopted.

Finally, substituting the RHSs of eqs. (4.191) and (4.192) in that of eq. (4.190)(4.133) yields, after some manipulation The function (4.136) can be minimized with respect to D by taking its partial derivative with respect to this variable and setting it to zero; this results in the equation that can be easily put in form

(4.195) where

(4.196) (4.197) and

(4.198)

It should be expected that only one of the two solutions of eq. (4.195) satisfies the inequalities (4.186)⁴ . If this solution is denoted D, the estimate (see eq. (4.185))

(4.199) of T can be computed.

4.1.5. Target cancellation procedure to be employed together with the single frequency estimator

A target cancellation procedure is used in combination with the SFE in T2 of the first embodiment. In this paragraph, we focus on the procedure expressed by eq. (2.15) and requiring the evaluation of the contribution given by the i- th (i.e., by the last) target detected on the v-th RVA. In our work, the computation of this contribution is based on the assumption that it is due to a point target for simplicity. Therefore, if denote the estimates of the frequency

⁴ In running our computer simulations, only the first-order problem has been solved. and the complex amplitude, respectively, characterizing this target, the expressions and are employed; here, denotes the iV₀-th order DFT of the vector

(4.203)

T 0,....,0 with k = 0, 1 and 2, the N₀-th order DFT of the vector

(4.204) and

(4.205) is the normalized frequency associated with the frequency

It is important to point out that an efficient method can be used for the computation of the vectors appearing in eqs. (4.200)-(4.202) (with k = 0, 1 and 2); note that, for any k, these vectors represent the iV₀-th order DFTs of the sequences and respectively, with In fact, the Z-th element of the vectors given by and respectively, where and

Therefore, the identities

and that hold for any p 6 ⁽C, can be exploited for an efficient computation of the RHSs of eqs. (4.206) and (4.207).

Finally, it is important to mention that our cancellation procedure allows to remove the contribution of a single target in each iteration of the STDREC algorithm. If a cluster of distinct frequencies is estimated by the SFE in the i-th iteration of the STDREC algorithm, each of the components of the triplet (Cj^ , C^_v) consists of the sum of terms and each term is evaluated on the basis of eqs. (4.200)-(4.202). The same considerations apply to the CSFE and the cancellation procedures illustrated in the following two paragraphs. 4.1.6. Target cancellation procedure to be employed with the complex single delay estimator - SFCW radar system

In this paragraph we take into consideration a MIMO SFCW radar system and focus on the two cancellation procedures employed in embodiment #1 of our method (see paragraph 2.3.1. Embodiment #1); the first procedure is used in combination with the CSDE in T2 (and, in particular, in STDREC-S2), whereas the second one is exploited in combination with the CSDE in T3-S1, and in particular in STDAEC-S2). The first procedure is expressed by eq. (2.15), that have been already illustrated for a MIMO FMCW radar system; however, the evaluation of the contributions (C®_v,

C®_j,, di^ven by the i-th (i.e., by the last) target detected on the v-th RVA is based on different formulas. These contributions, whose expressions are derived under the assumption that this target can be represented as a point target, are evaluated as

(4.214)

(4.215) and

(4.216) where is the estimate of the complex amplitude characterizing the considered target, denotes the N₀-th order IDFT of the vector

with k = 0, 1 and 2, and T^P is the estimate of the delay characterizing the considered target. It is important to point out that an efficient method can be used for the computation of the vector appearing in eqs. (4.214)-(4.216) (with k = 0, 1 and 2); note that, for any k, this vector is the iV₀-th order IDFTs of the sequence

— (0 n = 0,1,...,N — 1}, with W = W_v . In fact, the Z-th element of the vectors given by respectively, where Therefore, the identities (4.210)-(4.213) can be exploited for an efficient computation of the RHSs of eq. (4.220). The second cancellation procedure is expressed by eq. (2.57) and requires the evaluation of the contribution Cj^fZ] given by the i- th (i.e., by the last) target detected in the Z-th delay bin on the whole array. If A₍ \l\ , 7V_,i [Z] and T_{H i} [l] denote the estimates of the complex amplitude, the normalized vertical and horizontal spatial delay, respectively, characterizing this target, the expression with are employed for any RVA (i.e., for any p and q) . 4.1.7. Target cancellation procedure to be employed with the complex single frequency estimator - FMCW radar system

In this paragraph we take into consideration a MIMO FMCW radar system and focus on the two cancellation procedures employed in our method; the first procedure is used in combination with the CSFE in T2 (and, in particular, in STDREC-S2), whereas the second one is exploited in combination with the CSFE in T3-S1, and in particular in STDAEC-S2). The first procedure has been already illustrated for a MIMO FMCW radar system in combination with the SFE; however, in this case, the evaluation of the contributions given by the Z-th (i.e., by the last) target detected on the v-th RVA is based on different formulas, since the downconverted received signals made available by the considered radar system are complex. These contributions, whose expressions are derived under the assumption that the considered target can be represented as a point target, are evaluated as and here, W_{v k} denotes the N₀-th order DFT of the vector with k = 0, 1 and 2

(4.228) and

(4.229) is the normalized frequency associated with the frequency .

It is important to point out that the same efficient method described in Paragraph 4.1.5 can be used for the computation of the vector W_{v k} appearing in eqs. (4.200)-(4.202) (with Zc= 0, 1 and 2).

The second cancellation procedure requires the evaluation of the contribution C^f] given by the i-th (i.e., by the last) target detected in the Z-th frequency bin on the whole array; this target cancellation procedure is used in combination with the CSFE in T3 of both the embodiments described in the previous section for a MIMO FMCW radar system. Here, we focus on the target cancellation procedure employed in the first embodiment. This procedure is expressed by eq. (2.57) and requires the evaluation of the contribution cj^[], given by the Z-th (i.e., by the last) target detected in the Z-th frequency bin on the whole array; in our method, this target is represented as a point target for simplicity. Therefore, if Ai [l] , F_{V i} [l] and F_{H i} [Z] denote the estimates of the complex amplitude, the normalized vertical spatial frequency and the normalized horizontal spatial frequency, respectively, characterizing this target, the expression with is employed for any RVA (i.e., for any p and q).

4 .2 . Embodiment #2

The processing accomplished in the embodiment #2 of the proposed method is very similar to that of the first one, as illustrated in Paragraph 2.2.1. Embodiment #1. In fact, the only difference is represented by its third task (T3), since spatial estimation is accomplished serially i.e., on a bin-by-bin basis. In the following, we show how the cancellation procedure is accomplished in a MIMO FMCW radar system extracting the inphase component of the RF signals only, in a MIMO FMCW radar system extracting both the inphase and quadrature components of the RF signals and in a MIMO SFCW radar system.

4.2.1. Single bin cancellation - FMCW radar system In a MIMO FMCW radar system, the terms C®_v) appearing in eq. (2.68) are computed on the basis of eqs. (4.200)-(4.202) if the received sequence {r_vn} is real and eqs. (4.224)-(4.226) if the received sequence {r_vn} is complex; however, in this case, the presence of D[l] distinct targets detected in the ct_j-th frequency bin and the fact that the frequencies of those targets do not necessarily fall in that bin must be taken into account. In practice, the formulas and are used in the case of inphase component only, and and in the case of inphase and quadrature components; in both cases, v = 0, 1, ..., N_VR — 1 . In the last six equations, [Z] denotes the complex amplitude associated with the i-th target detected in

- the a₍-th frequency bin, and the complex vectors are computed on the basis of eqs . (4.206)-(4.207), where, however, the normalized frequency is used in place of (4.205) in the computation of the quantity (4.204); moreover, /. [Z] represents the frequency associated with the i-th target detected in the ct_j-th frequency bin.

4.2.2. Single bin cancellation - SFCW radar system

In a MIMO SFCW radar system, the evaluation of the terms C®_v) employed in the SBC procedure of embodiment #2 (see eq. (2.68)) is based on some formulas different from the ones employed in a MIMO FMCW radar system. In practice, the formulas

and are used for Here, D[l] represents the overall number of distinct targets detected in the a₍-th time bin and Ai [Z] denotes the complex amplitude associated with the i-th target detected in the a₍-th time bin. Moreover, the complex vectors are computed on the basis of eqs. (4.220) (with k = 0,1 and 2), where, however, the normalized delay (see eq. (1.27)) is used in place of (see eq.

(4.219)) in the computation of the quantity (see eq. (4.218)), being an estimate of the delay associated with the i-th target detected in the a₍-th time bin.

Finally, it is worth noting that eqs. (4.238)-(4.240) are structurally similar to eqs. (4.232)-(4.234) derived for a MIMO FMCW radar system. 5. Novelty of the Developed Estimation Techniques and Other Potential Applications of Them

The iterative estimation technique called single target detection, range estimation and cancellation (STDREC) and employed in task 2, step 2, of both the proposed embodiments for a FMCW radar system solves the well known problem of estimating the parameters of multiple undamped sinusoids; for this reason, it can be potentially used in any application, not necessarily related to the field of radar systems, in which other estimators solving the same problem are adopted. It is important to stress that the problem of estimating the parameters (and, in particular, the frequencies) of multiple sinusoids in white noise is a classic one and has been deeply investigated in the technical literature. It has been shown that the periodogram method represents a realization of the maximum likelihood (ML) estimator if the spacing among the normalised frequencies of all the sinusoids is greater than 1/iV, where N is the data record length (see "D. C. Rife and R. R. Boorstyn, "Multiple tone parameter estimation from discrete-time observations," The Bell System Technical Journal, vol. 55, no. 9, pp. 1389-1410, 1976."). If, however, this condition is not met, ML estimation leads to a nonlinear least-squares problem. Therefore, in this case, the implementation of the ML estimator requires a nonlinear minimization which, generally speaking, is not guaranteed to converge or may converge to only a local minimum; in addition, its computational burden is high. This explains why a number of suboptimal solutions have been proposed in the technical literature. Most of them are based on iterative algorithms (e.g., see "S. Kay, "Accurate frequency estimation at low signal- to-noise ratio," IEEE Trans. Ac., Speech and Sig. Proc., vol. 32, no. 3, pp. 540-547, 1984.") and usually exploit a two stage approach; in the first stage, suboptimal initial estimates of the parameters of interest are obtained, whereas in the second one such estimates are refined through an iterative maximization of the likelihood function (e.g., see "P. Stoica, R. L. Moses, B. Friedlander, and T. Soderstrom, "Maximum likelihood estimation of the parameters of multiple sinusoids from noisy measurements," IEEE Trans. Ac., Speech, and Sig. Proc., vol. 37, no. 3, pp. 378-392, 1989." and references therein). The novelty of our solution is mainly represented by the SFE technique and by the simplicity of the adopted cancellation procedure. Computer simulations have evidenced that: a) the SFE technique is able to generate an accurate estimate of the parameters of each sinusoid through the computation of both the integer part and its fractional part of its frequency (see eq. (2.12)); b) the accuracy of frequency estimation plays a fundamental role in the efficacy of the adopted serial cancellation procedure, since it mitigates the error accumulation phenomenon occurring in any algorithm implementing the concept of estimation of multiple tones through serial cancellation. Similar considerations apply to the CSFE, which is used to solve the well known problem of estimating the parameters of multiple undamped complex exponentials; for this reason, it can be potentially used in any application, not necessarily related to the field of radar systems, in which other estimators solving the same problem are adopted.

If the STDREC algorithm is employed in the RPE of a SFCW radar system, an estimate of the channel impulse response (CIR) is obtained for each of the selected N_A RVAs. Moreover, such a CIR is represented by a sequence of lines (i.e., of delayed delta functions), each positioned in a specific time bin⁵ and characterized by a specific complex gain. It is important to note that the received signal model (1.23) has the same mathematical structure as the signal available at the output of the IDFT evaluated by the demodulator in a digital communication system employing the orthogonal frequency division multiplexing (OFDM) modulation (e.g., see eq. (4.129) in Par. 4.4.4 of "G. M. Vitetta, F. Pancaldi, D. P. Taylor, and P. Martin, Wireless Communications: Algorithmic Techniques. John Wiley & Sons.") when known (i.e., pilot) channel symbols are transmitted. For this reason, the STDREC algorithm can be exploited for accomplishing pilot-based channel estimation of the impulse response of the communication channel in this system. This result has significant technical implications, since the OFDM has been adopted in the air interface of fourth generation (4G) and fifth generation (5G) standards for cellular communications (see "A. A. Zaidi, R. Baldemair, V. Moles-Cases, N. He, K. Werner, and A. Cedergren, "Ofdm numerology design for 5g new radio to support iot, embb, and mbsfn," IEEE Communications Standards Magazine, vol. 2, no. 2, pp. 78-83, 2018." and "S. Srikanth, P.A. Murugesa Pandian, and X. Fernando, "Orthogonal frequency division multiple access in wimax and lte: a comparison," IEEE

Communications Magazine, vol. 50, no. 9, pp. 153-161, 2012.").

In fact, in any wireless link of a cellular communication system employing the OFDM modulation, an accurate knowledge of the CIR is required to compensate for the channel distortion at the receive side, i.e. for channel equalization (e.g., see Sec. 6.3 of "G. M. Vitetta, F. Pancaldi, D. P. Taylor, and P. Martin,

⁵ This represents the dual of the frequency bin defined in Paragraph 2.3. SFCW radar system. Wireless Communications: Algorithmic Techniques. John Wiley & Sons."). Moreover, the knowledge of the delay characterizing the first sample of the CIR (i.e., characterizing the first received echo of the transmitted signal) can be exploited to estimate the time-of-flight (TOF) of the signal travelling from a base station to a mobile phone or in the opposite direction, and these information can be exploited for precise positioning of cellular phones (see "F. Gustafsson and F. Gunnarsson, "Mobile positioning using wireless networks: possibilities and fundamental limitations based on available wireless network measurements," IEEE Signal Processing Magazine, vol. 22, no. 4, pp. 41-53, 2005."). It is worth mentioning that, as far as we know, the use of serial cancellation approach to channel estimation has been proposed in Par. II-A of ref. "X. Wang and S. B. Wicker, "Channel estimation and feedback with continuous time domain parameters," in Proc. 2013 IEEE Global Communications Conference (GLOBECOM), 2013, pp. 4306-4312." for a long term evolution-advanced (LTE-advanced) cellular system; however, the estimation technique developed in that manuscript is completely different from the STDREC technique we have devised for a SFCW radar system.

The iterative estimation algorithm called single target detection, angular estimation and cancellation (STDAEC) and employed in task 3, step 1, of both the proposed embodiments for a FMCW radar system solves the well known problem of estimating the parameters of multiple undamped complex exponentials. For this reason, it can be potentially used in any application, not necessarily related to the field of radar systems, in which other estimators solving the same problem are adopted. It is worth mentioning that the estimation of the parameters of multiple superimposed exponential signals in additive Gaussian noise represents a fundamental problem in time series analysis and system identification, and in antenna array processing. Various high resolution methods achieve excellent performance in terms of estimation accuracy at high SNR and with long data records. However, the performance of these methods is severely degraded when the SNR and/or data record length fall below some threshold, which may be significantly higher than that of the optimal ML estimate. This behavior may originate from the fact that the above methods are essentially heuristic modifications of algorithms yielding exact results when there is no noise, or when the amount of available data is infinite. Unluckily, direct maximization of the likelihood function, which is highly nonlinear in the unknown signal parameters, requires a computationally expensive multidimensional search. This has motivated the development of suboptimal iterative methods able to solve this difficult problem (e.g., see "J. Selva, "ML Estimation and Detection of Multiple Frequencies Through Periodogram Estimate Refinement," IEEE Sig. Proc. Lett., vol. 24, no. 3, pp. 249-253, 2017.", "Y. Bresler and A. Macovski, "Exact maximum likelihood parameter estimation of superimposed exponential signals in noise," IEEE Transactions on Acoustics, Speech, and Signal Processing, vol. 34, no. 5, pp. 1081-1089, 1986.", and "L. S. Biradar and V. U. Reddy, "A newton-based ziskind-wax alternating projection algorithm applied to spectral estimation," IEEE Transactions on Signal Processing, vol. 41, no. 3, pp. 1435-1438, 1993." and references therein). An interesting example of this approach is offered by the estimation technique developed in "J. Selva, "ML Estimation and Detection of Multiple Frequencies Through Periodogram Estimate Refinement," IEEE Sig. Proc. Lett., vol. 24, no. 3, pp. 249-253, 2017.". In each iteration of this algorithm, two steps are executed; in the first step, a single frequency or a frequency cluster is detected and estimated using the periodogram, whereas in the second one this estimate/these estimates are refined through a Newton-type method; moreover, once a set of exponentials is detected and estimated, its contribution is subtracted from the log-likelihood function. These two steps are iteratively repeated until no frequency whose contribution to the likelihood function is above a fixed detection threshold is found. The overall structure of this algorithm is conceptually similar to that of the STDAEC algorithm. However, unlike the STDAEC algorithm, it does not make any attempt to estimate the fractional part of each frequency through multiple FFTs; moreover, the cancellation of the contribution of each detected and estimated complex exponential is based on a completely different procedure (see eq. (12) in Sec. Ill of "J. Selva, "ML Estimation and Detection of Multiple Frequencies Through Periodogram Estimate Refinement," IEEE Sig. Proc. Lett., vol. 24, no. 3, pp. 249-253, 2017.").

As already explained above, the STDREC technique developed for a SFCW radar system can be employed in an OFDM communication system for pilot-based CIR estimation. In fact, the received signal model (1.23) has the same mathematical structure as the signal available at the output of the IDFT evaluated by the demodulator in a digital communication system employing the orthogonal frequency division multiplexing (OFDM) modulation (e.g., see eq. (4.129) in Par. 4.4.4 of "G. M. Vitetta, F. Pancaldi, D. P. Taylor, and P. Martin, Wireless Communications: Algorithmic Techniques. John Wiley & Sons.") when known (i.e., pilot) channel symbols are transmitted.

Similarly, the STDAEC technique employed in a SFCW radar system can be used in a MIMO OFDM communication system for pilot-based estimation of the direction of arrival (DOA) of an OFDM signal. As already mentioned above, OFDM represents the digital modulation format employed in the downlink of 5G cellular systems. In such systems, the STDAEC technique can be used in any base station equipped with a receive array to estimate the DOA of the signals transmitted by a mobile terminal served by the base station itself.

Finally, it is worth mentioning that the exploitation of multiple FFTs in the estimation of the frequency of a complex exponential has been previously proposed in "E. Chiavaccini and G. M. Vitetta, "Maximum-likelihood frequency recovery for ofdm signals transmitted over multipath fading channels," IEEE Trans. Commun., vol. 52, no. 2, pp. 244-251, 2004." and "F.Zuccardi Merli and G. M. Vitetta, "Blind feedforward frequency estimation for OFDM signals transmitted over multipath fading channels," IEEE Transactions on Wireless Communications, vol. 6, no. 6, pp. 2055-2059, 2007" to solve the frequency estimation problem at the receive side of a digital communication system employing the OFDM digital modulation. However, the algorithms developed for that applications are completely different from that employed by the CSFE.

Therefore, the present invention can be easily implemented in the field of 5G base-station, the point targets are replaced by the mobile terminals and the virtual antennas are replaced by the receiving antennas of the 5G base station.

It is apparent to the skilled person in the art that all the equations described above can rewritten without modifications, simply considering the above slight differences. 6. Numerical Results

In this section, some numerical results, generated by means of computer simulations in three different scenarios, are illustrated. The first scenario aims at assessing the accuracy provided by the STDREC algorithm in a single-input single-output (SISO) FMCW radar system. The second scenario allows us to show that the STDAEC algorithm is able to detect multiple targes contributing to the same frequency bin in a MIMO FMCW radar system and to generate accurate estimates of their spatial coordinates. Finally, the third scenario aims at showing that the accuracy of the 2D and 3D images generated by the two proposed embodiments is substantially better than that provided by standard algorithms based on the computation of multidimensional FFTs; in this case, both FMCW and SFCW radar systems are considered.

In all the considered scenarios, estimation accuracy is assessed by evaluating the root mean square error (RMSE) and the peak error

(6.2) here, Xi and represent the exact value of the parameter to be estimated for the Z-th target (with Z= 0,1, L— 1) and the corresponding estimate. The numerical results illustrated below refer to three different simulation scenarios.

Scenario #1 - The first scenario is characterized by four point targets having different radar cross sections (RCSs). The exact range and RCS of each of these targets are listed in Table 1. An FMCW radar system equipped with a single transmit and a single receive antenna is considered in this case and the STDREC algorithm is employed for range estimation only. Moreover, the following parameters have been selected for the FMCW system: m = 1.5-10¹³ Hz/s, B = 3 GHz (see eq. (1.4)), f₀ = 77 GHz, f_s = 5 MHz and N = 1024; moreover, the variance of the noise affecting the received signal sample r_vn (1.12) is = 3.0 (corresponding to a SNR of about 3.5 dB). As far as the STDREC parameters are concerned, the following choices have been made: M = 8 (see (2.7)), iV_SFE = 5 and the threshold TSTDREC is set to 10% of the energy of the vector X₀ (see eq. (2.16)). The amplitude spectrum of the received signal and the residual spectrum available at the end of the cancellation procedure accomplished by the STDREC are illustrated in Figure 17; the estimates of the ranges of the targets and of their RCSs are listed in Table 1. These results lead to the conclusion that that our method is able to achieve good accuracy both in terms of range and RCS estimation. Note also that the accuracy achieved in range estimation is better than that obtained in RCS estimation; in fact, from Table 1 it is easily inferred that e_R = 0.003 m, e_R = 0.006 m, a_RCS = 0.05 V and £RCS= 0.08 V (see eqs. (6.1)—(6.2)).

Scenario #2 - The second scenario is characterized by four point targets having the same range and placed at the vertices of a rectangle; moreover, the line connecting the centre of the radar with that of the rectangle is orthogonal to the rectangle itself. A 3D representation of the targets together with the estimated position and the associated RCS is shown in Figure 18. The coordinates and the RCS of each of these targets are listed in Table 2. In this case, our simulation results refer to a MIMO FMCW radar system equipped with the antenna array described in Paragraph 3.2 and characterized by the following parameters: m = 1.5·10¹³ Hz/s, B = 3 GHz (see eq. (1.4)), f₀ = 77 GHz, f_s = 5 MHz, N = 1024. Moreover, the variance of the noise affecting the received signal sample r_vn (1.12) is = 0.2 (the corresponding SNR is about 22.0 dB). As far as the parameters of the STDREC algorithm are concerned, the following choices have been made: ^ = 10, M = 8 (2.7) and iV_SFE = 5; moreover, the threshold T_STDRE_C is equal to 80% of the initial energy of the vector X₀ (see eq. (2.16)). In this case, the STDREC algorithm allows to detect the presence of at least one target in the frequency bin b₀ = 493. The estimation of the azimuth and of the elevation of the four targets contributing to this bin is accomplished by the STDAEC algorithm; the following choices have been made for its parameters: iV₀ = 256 (see eq. (2.37)), iV_Vu_LA ⁼ 16, iV_HULA ⁼ 31, P_R = 16, q_R = 8, iV_CSFE = 5; moreover, the threshold T_STDAE_C I^S equal to the 20% of the initial energy measured in the considered frequency bin (see eq. (2.58)).

The amplitudes of the elements of the vector X_yu_LA^] (see eq.

(2.33)), the amplitudes of the elements of the vector X_Q ^VF^[0] generated by vertical folding (see eq. (2.45)) and the spatial frequencies estimated for the first target of Table 2 are shown in Figure 19. In this case, the STDAEC algorithm allows to detect all the targets, and to estimate their horizontal and vertical spatial frequencies. The estimates of the spatial coordinates of these targets are listed in Table 2; in this case, s_R = 0.011 m, e_R = 0.015 m, e_q = 0.9°, e_q = 0.9°, ey = 0.1°, 0.1° and e_RCS = 0.07 V, ^R_CS ⁼ 0.12 V (see eqs. (6.1)-(6.2)). These results lead to the conclusion that the joint use of the STDREC and the STDAEC algorithms works well, even in the presence of multiple targets contributing to the same frequency bin.

Scenario #3 - The third scenario is characterized by nine point targets grouped in three clusters, each made of three points having similar ranges. The target coordinates, the RCS and their estimates for a FMCW (SFCW) radar system are listed in Table 3 (Table 5) for the first embodiment and in Table 4 (Table 6) for the second embodiment.

As far as the FMCW radar system is concerned, our simulation results are obtained under the following assumptions. The RF signal is radiated by a radar system equipped with the antenna array described in Paragraph 3.2 and characterized by the following parameters: m = 1.5 · 10¹³ Hz/s, B = 3 GHz (see eq. (1.4)), /o = 77 GHz, f_s = 5 MHz, N = 1024; moreover, the variance of the noise affecting the received signal sample r_vn (1.12) is = 0.2 (the corresponding SNR is about 22.0 dB).As far as the parameters of the STDREC algorithm are concerned, the following choices have been made: iV_A = 10, M = 8 (2.7) and iV_SFE = 5 ; moreover, the threshold 7s_TDREC is equal to 4% of the energy ||X_{0 Vfe}|| ^on the v_k- th RVA. The estimation of the azimuth and of the elevation of all the targets is accomplished by the STDAEC algorithm; the following choices have been made for its parameters: N₀ = 256 (see eq. (2.37)), iV_VULA ⁼ 16, ^VHULA ⁼ 31, P_R ⁼ 16, q_R = 8, iV_CSFE = 5; moreover, the threshold T_STDAEC is equal to the 3% of the energy measured in the considered frequency bin (see eq.

(2.30)). The SFCW radar system, instead, is equipped with an URA (like the one illustrated in Figure 9) composed by N_vv = 20 antennas along the vertical dimension and N_VH = 16 physical elements along the horizontal axis. The overall number of virtual channels is N_VR = 20x 16 = 320. This device is characterized by the following parameters: B = 3 GHz (see eq. (1.20)), f₀ = 77 GHz, F_s = 3 MHz, N = 1024; moreover, the variance of the noise affecting the frequency response sample H_vn (1.23) is = 0.2 (the corresponding SNR is about 22.0 dB). As far as the parameters of the STDREC algorithm are concerned, the following choices have been made: iV_A = 10, M = 8 (2.7) and iV_SDE = 5; moreover, the threshold T_STDRE_C is equal to

4% of the energy ||X_0Vfe|| ^on the i¾-th RVA. The estimation of the azimuth and of the elevation of the nine targets is accomplished by the STDAEC algorithm; the following choices have been made for its parameters: N₀ = 320 (see eq. (2.37)), iV_VULA = 20, iV_HULA = 16, p_R = 8, q_R = 10, iV_CSDE = 5 ; moreover, the threshold T_STDAE_C is equal to the 3% of the initial energy ||X®[/]|| measured in the considered frequency bin (see eq. (2.30)). The 3D image generated by both radar systems through the first embodiment is shown in Figure 20. In this case, the STDAEC algorithm allows to detect all the targets, and to estimate their horizontal and vertical spatial frequencies. Moreover, the RMSE and the peak error achieved by both systems are listed in Table 7. These results lead to the conclusion that the joint use of the STDREC and the STDAEC algorithms works well, even in the presence of multiple extended targets.

Tables:

Claims

1. [SPE-2D] Method for estimating the angular coordinates of point targets whose range has been estimated through a MIMO FMCW radar equipped with a plurality of transmitting (TX) and receiving (RX) antennas, wherein each couple of said TX and RX antennas is replaced with the equivalent virtual antenna; said virtual antennas form a virtual array composed by one or multiple horizontal uniform linear arrays (HULAs); a receive side of the MIMO FMCW radar being arranged to generate real or complex down- converted signals in response to a propagation scenario including a plurality of detected point targets; said method includes:

— (v)

(Step 7) acquisition of a spectrum (X₀ ) of said signal and its first N_FFT — 1 derivatives when said signal is real and N_FFT = 4 when said signal is complex, for each of said virtual antennas,

(Step 8) acquisition of a list of a predetermined number of discretized ranges, named frequency bins, of said detected point targets, - (Step 9) selection of a reference antenna, a reference horizontal uniform linear array, named Reference HULA, from said virtual array,

(Step 10 [STDAEC]) sequential execution for each of said frequency bins; said execution includes: - (Step 10.1 [STDAEC-S1]) estimation of the horizontal spatial frequency and complex amplitude of the most dominant point target; said estimation includes: - (Step 10.1.4 [STDAE-S4]) computation of the spectrum {s_HULA0) and its first three derivatives ({^s _HULAk> k = 1,2,3}) of the sequence collecting the samples of said spectrum (X₀) referred to said Reference HULA in the considered frequency bin; said computations being carried out through FFT calculation with a second oversampling factor (M),

- (Step 10.1.5 [STDAE-S4]) estimation of said horizontal spatial frequency,

- (Step 10.1.6 [STDAE-S5]) checking whether combination of the spectra associated with said Reference HULA, after compensating for their phase differences on the basis of said estimated horizontal frequency, generates an amplitude peak in the absolute value of the resulting spectrum, named overall folded spectrum; and when the checking is positive execution of the following steps:

- (Step 10.1.7) refinement of said range and complex amplitude of said most dominant point target;

(Step 10.2 [STDAEC-S2]) cancellation of said most dominant point target,

(Step 10.3 [STDAEC-S3]) calculation of a residual energy in the considered frequency bin and its comparison with a predetermined threshold (T_STDAEC) , and checking whether said residual energy exceeds said predetermined threshold and, if so, return to computation of the spectrum (Step 10.1.4) and go to computation of the spatial coordinates (Step 11), otherwise, in negative case of said checking (Step 10.1.6) execution of the following steps: - (Step 10.4) updating of said acquired list of said frequency bins and return to Step 10 for a further frequency bin or exit when all frequency bins are examined, wherein said sequential execution (Step 10) ends when all the frequency bins have been analyzed, then

(Step 11) computation of the spatial coordinates of all the point targets belonging to all the frequency bins on the basis of said spatial frequencies and said complex amplitude and generation of an overall image of the propagation scenario.

2. [SPE-3D] Method according to claim 1, wherein said virtual antennas form a virtual array composed by one or multiple horizontal uniform linear arrays (HULAs) and one or multiple vertical uniform linear arrays (VULAs); the method including the following steps:

— (v)

- said (Step 7) acquisition of a spectrum (X₀ ) of said signal

— (v) and its first N_FFT — 1 derivatives ({X_fc ;k = 1,...,N_FFT 1}) with N_FFT = 3 when said signal is real and N_FFT = 4 when said signal is complex, for each of said virtual antennas, - said (Step 8) acquisition of a list of a predetermined number of discretized ranges, named frequency bins, of said detected point targets,

- said (Step 9) selection of a reference antenna, and in addition a reference vertical uniform linear array, named Reference VULA, a reference horizontal uniform linear array, named Reference HULA, and in addition, at least a sub-set with a number of said horizontal uniform linear arrays (HULAs) different from the Reference HULA, from said virtual array, - said (Step 10 [STDAEC]) sequential execution for each of said frequency bins; said execution includes: said (Step 10.1 [STDAEC-S1]) estimation of the horizontal spatial frequency and complex amplitude of the most dominant point target and in addition estimation of the vertical spatial frequency; said estimation includes:

- (Step 10.1.1 [STDAE-SI]) computation of the spectrum (s_VULA0) of the sequence collecting the samples of said spectrum (X₀) referred to said Reference VULA in the considered frequency bin and its first 3 derivatives 1,2,3}) through FFT calculation with a first oversampling factor (M),

- (Step 10.1.2 [STDAE-S2]) estimation of said vertical spatial frequency,

- (Step 10.1.3 [STDAE-S3]) combination of the spectra associated with the Reference HULA and said sub-set of HULAs, after compensating for the phase differences between the Reference HULA and said sub-set of HULAs employing said estimate of the vertical spatial frequency (F_v) ; said superposition, named vertical folding, generates a vertically folded spectrum, said (Step 10.1.4 [STDAE-S4]) computation of the spectrum (^SHULA_,o) and its first three derivatives 1,2,3}) of the sequence collecting the samples of said spectrum (X₀) referred to said Reference HULA in the considered frequency bin, and said spectrum is referred to said vertically folded spectrum; said computations being carried out through FFT calculation with a second oversampling factor (M),

- said (Step 10.1.5 [STDAE-S4]) estimation of said horizontal spatial frequency, - (Step 10.1.6 [STDAE-S5]) checking whether combination of the spectra associated with said Reference HULA, after compensating for their phase differences on the basis of said estimated horizontal frequency and of said estimated vertical frequency, generates an amplitude peak in the absolute value of the resulting spectrum, named overall folded spectrum, wherein the spectra are associated with said virtual antennas of said subset of HULAs; and when the checking is positive execution of the following steps: said (Step 10.1.7) refinement of said range and complex amplitude of said most dominant point target;

- said (Step 10.2 [STDAEC-S2]) cancellation of said most dominant point target,

- said (Step 10.3 [STDAEC-S3]) calculation of a residual energy in the considered frequency bin and its comparison with a predetermined threshold {TSTDAEC)_t and checking whether said residual energy exceeds said predetermined threshold and, if so, return to said computation of the spectrum (Step 10.1.1), otherwise go to computation of the spatial coordinates (Step 11), otherwise, in negative case of said checking (Step 10.1.6) execution of the following steps: said (Step 10.4) updating of said acquired list of said frequency bins and return to Step 10 for a further frequency bin or exit when all frequency bins are examined, wherein said sequential execution (Step 10) ends when all the frequency bins have been analyzed, then - said (Step 11) computation of the spatial coordinates of all the point targets belonging to all the frequency bins on the basis of said spatial frequencies and said complex amplitude and generation of an overall image of the propagation scenario.

3. Method according to claim 1 or 2, wherein said execution (Step 10- [EMB#1]) is carried out by running multiple parallel processes whose number is equal to the number of said frequency bins of said acquired list and said generation of the overall image (Step 11) is carried out when all parallel processes have been executed.

4. Method according to claim 1 or 2, wherein said acquisition (Step 8 - [EMB#2]) includes the acquisition of target energy and wherein said execution (Step 10 - [EMB#2]) is repeated sequentially over the acquired list, ordered according to said target energy, and wherein said generation of the overall image (Step 11) is carried out when all the frequency bins have been analyzed.

5. Method according to claim 1 or 2, wherein said (Step 8) acquisition includes:

- (Step 8.1) acquisition of a list of a predetermined number of discretized ranges of said plurality of point targets by selecting the indices of distinct frequency bins where at least one point target is detected.

6. [STDAEC-S1] Method according to claim 1, wherein said estimation (Step 10.1) includes a preliminary step (Step 10.1.0) including: a) setting the iteration index i to 1,

— (v) b) defining a matrix collecting said spectral information (X₀ ) available for each of said virtual antennas and referring to the considered frequency bin only.

7. [STDAEC-S1] Method according to claim 2, wherein said estimation (Step 10.1) includes a preliminary step (Step 10.1.0) including: a) setting the iteration index i to 1,

— (v) b) defining a matrix collecting said spectral information (X₀ ) available for each of said virtual antennas and referring to the considered frequency bin only c) extracting a vector, referring to said Reference VULA only, from said matrix.

8. [STDAE-S2] Method according to claim 2, wherein said estimation of said vertical spatial frequency (Step 10.1.2) includes

- (Step 10.1.2.1.0) representation of a target vertical spatial frequency (F_v) , as sum of a main portion (P_V^_OFT) ^and ^a residual portion (6_VF_dft) , where F_DFT = l/iV₀, where N₀ is the FFT order, wherein said main portion is determined by searching the index (b_n) of the vector element s_VULAOp_v corresponding to the maximum absolute value of the same vector s_VULA0 and said residual frequency computation is found through an iterative procedure together with the corresponding target complex amplitude A = aexp(Jxp), where a is said target amplitude and y is said target phase.

9. [STDAE-S4] Method according to claim 1 or 2, wherein said horizontal spatial frequency estimation (Step 10.1.5) includes:

(Step 10.1.5.1.0) representation of a target horizontal spatial frequency (F_H), as sum of a main portion (bp^rt) and a residual portion (6_HF_dft), where F_DFT = 1/N_0r where N₀ is the FFT order, wherein said main portion is determined by searching the index (b_H) of the vector element s_HULAOp_H corresponding to the maximum absolute value of the same vector s_HULA0 and said residual frequency computation is found through an iterative procedure together with the corresponding target complex amplitude A = aexp(ji|/), where a is said target amplitude and y is said target phase.

10. [STDAE-S5] Method according to claim 1, wherein said estimate of the horizontal spatial frequency (F_H) is employed to compute the phase difference between said reference antenna and said virtual antennas of said Reference HULA to be used in said constructive combination (Step 10.1.6) to compute said overall folded spectrum (X_0F,O) and its first three derivatives ({X_0Ffc;k =

1,2,3}) by combining the spectral contributions referred to said sub-set of virtual antennas.

11. [STDAE-S5] Method according to claim 2, wherein said estimate of the vertical spatial frequency (F_v) and said estimate of the horizontal spatial frequency (F_H ) are employed to compute the phase difference between said reference antenna and said virtual antennas of said sub-set of HULAs to be used in said constructive combination (Step 10.1.6) to compute said overall folded spectrum and its first three derivatives ) by combining the spectral contributions referred to said subset of virtual antennas.

12. Method according to claim 1 or 2 wherein said refinement (Step 10.1.7) includes:

(Step 10.1.7.1.0) representation of a target normalized frequency where f_s is the sampling frequency, as sum of a main portion (aF_dft) and a residual portion (5F_dft), where F_DFT = l/iV₀, where iV₀ is the FFT order, wherein said main portion is determined by searching the index (a) of the vector element X_OF,o_,a corresponding to the element of the vector X_OF,O having the maximum absolute value and said residual frequency is found through an iterative procedure together with the corresponding target complex amplitude A = aexp(jxp), where a is said target amplitude and y is said target phase;

13. [STDAEC-S2-EMB#1] Method according to claim 1 or 2 wherein said cancellation (Step 10.2) includes the computation and subtraction of the contribution given by said dominant point target to the spectral contribution referred to said considered frequency bin to generate a new residual of the spectrum and of the corresponding derivatives.

14. [STDAEC-S2-EMB#2] Method according to claim 1 or 2 wherein said cancellation (Step 10.2) includes the computation and subtraction of the contribution given by said dominant point target to said spectrum and to the corresponding derivatives to generate a new residual of the spectrum and of the corresponding derivatives.

15. Method according to claim 1, wherein said computation (Step 11) includes the computation of the range (R) , and the azimuth (Q) as: and respectively, here, m is the radar chirp rate, c is the speed of light, l is the radar wavelength, d_VH is the distance between two adjacent virtual antennas of said HULA, f_s is the sampling frequency, F is said discretized range, and F_H is said horizontal spatial frequency.

16. Method according to claim 2, wherein said computation (Step

11) includes the computation of the range the elevation and the azimuth as: and respectively, here, m is the radar chirp rate, c is the speed of light, l is the radar wavelength, d_vv is the distance between two adjacent virtual antennas of said VULAs and d_VH is the distance between two adjacent virtual antennas of said HULAs, f_s is the sampling frequency, F is said discretized range, F_v is said vertical spatial frequency, and F_H is said horizontal spatial frequency.

17. [RPE] Method according any one of the previous claims further comprising the following preliminary steps:

- (Step 1: FFT processing) estimation of said spectrum ( of said signal and its first N_FFT — 1 derivatives 1}) with N_FFT = 3 when said signal is real and N_FFT = 4 when said signal is complex, for each of said virtual antennas through FFT calculation with an oversampling factor ( M ) ,

- (Step 2: RVA selection) acquisition of a sub-set of a number (N_a) of virtual antennas spectra and of the corresponding derivatives,

Recursive execution of

(Step 3: STDREC-S1) calculation of an estimate of the parameters of the most dominant point target, including phase (y) , amplitude ( a) and frequency (/) for each of said virtual antennas of said sub-set on the basis of said spectrum, and its derivatives, through FFTs for each of said virtual antennas of said sub-set,

- (Step 4: STDREC-S2) cancellation of said most dominant point target and computation of a residual spectrum and its first

1 derivatives,

- (Step 5: STDREC-S3) computation of an energy of said residual spectrum and return to (Step 3) until said energy is over a predetermined threshold otherwise - (Step 6) computation of said average energy and fusion of ranges information of said point targets detected by said number of virtual antennas to list said point targets according to said average energy in a decreasing order.

18. [SPE-2D] Method for estimating the angular coordinates of point targets whose range has been estimated through a MIMO SFCW radar equipped with a plurality of transmitting (TX) and receiving (RX) antennas, wherein each couple of said TX and RX antennas is replaced with the equivalent virtual antenna; said virtual antennas forming a virtual array composed by one or multiple horizontal uniform linear arrays (HULAs); a receive side of said MIMO SFCW radar is arranged to generate complex down-converted signals in response to a propagation scenario including a plurality of detected point targets; said method includes:

(Step 7) acquisition of an impulse response characterizing the communication channel associated with the said equivalent virtual antenna and its first derivatives with for each of said virtual antennas,

(Step 8) acquisition of a list of a predetermined number of discretized ranges, named delay bins, of said detected point targets,

(Step 9) selection of a reference antenna, a reference horizontal uniform linear array, named Reference HULA, from said virtual array,

- (Step 10 [STDAEC]) sequential execution for each of said delay bins; said execution includes:

- (Step 10.1 [STDAEC-S1]) estimation of the horizontal spatial delay and complex amplitude of the most dominant point target said estimation includes:

- (Step 10.1.4 [STDAE-S4]) computation of the spectrum and its first three derivatives ) of the sequence collecting the samples of said impulse response ) referred to said Reference HULA in the considered delay bin,; said computations are carried out through IFFT calculation with an oversampling factor

- (Step 10.1.5 [STDAE-S4]) estimation of said horizontal spatial delay,

- (Step 10.1.6 [STDAE-S5]) checking whether combination of the impulse responses associated with said virtual antennas of said Reference HULA, after compensating for their phase differences on the basis of said estimated horizontal delay, generates an amplitude peak in the absolute value of the resulting impulse response, named overall folded impulse response; and when the checking is positive execution of the following steps:

- (Step 10.1.7) refinement of said range and complex amplitude of said most dominant point target; - (Step 10.2 [STDAEC-S2]) cancellation of said most dominant point target,

(Step 10.3 [STDAEC-S3]) calculation of a residual energy in the considered delay bin and its comparison with a predetermined threshold and checking whether said residual energy exceeds said predetermined threshold and, if so, return to computation of the spectrum (Step 10.1.4), otherwise go to computation of the spatial coordinates (Step 11) otherwise, in negative case of said checking (Step 10.1.6) execution of the following steps: - (Step 10.4) updating of said acquired list of said delay bins and return to Step 10 for a further delay bin or exit when all delay bins are examined, wherein said execution (Step 10) ends when all the delay bins have been analyzed, then - (Step 11) computation of the spatial coordinates of all the point targets belonging to all the delay bins on the basis of said spatial delays and said complex amplitude and generation of an overall image of the propagation scenario.

19. [SPE-3D] Method according to claim 18, wherein said virtual antennas form a virtual array composed by one or multiple horizontal uniform linear arrays (HULAs) and one or multiple vertical uniform linear arrays (VULAs); the method including the following steps: said (Step 7) acquisition of an impulse response characterizing the communication channel associated with the said equivalent virtual antenna and its first derivatives for each of said virtual antennas,

- said (Step 8) acquisition of a list of a predetermined number of discretized ranges, named delay bins, of said detected point targets,

- said (Step 9) selection of a reference antenna, a reference vertical uniform linear array, named Reference VULA, a reference horizontal uniform linear array, named Reference HULA, and at least a sub-set with a number of said horizontal uniform linear arrays (HULAs) different from the Reference HULA from said virtual array,

- said (Step 10 [STDAEC]) sequential execution for each of said delay bins; said execution includes: said (Step 10.1 [STDAEC-S1]) estimation of the horizontal spatial delay and complex amplitude of the most dominant point target and estimation of the vertical spatial delay; said estimation includes:

- (Step 10.1.1 [STDAE-SI]) computation of the spectrum of the sequence collecting the samples of said impulse response referred to said Reference VULA in the considered delay bin and its first 3 derivatives through IFFT calculation with an oversampling factor - (Step 10.1.2 [STDAE-S2]) estimation of said vertical spatial delay,

- (Step 10.1.3 [STDAE-S3]) combination of the spectra associated with the Reference HULA and said sub-set of HULAs, after compensating for the phase differences between the Reference HULA and said sub-set of HULAs employing said estimate of the vertical spatial delay (T_v) ; said superposition, named vertical folding, generates a vertically folded impulse response, said (Step 10.1.4 [STDAE-S4]) computation of the spectrum and its first three derivatives ( ) of the sequence collecting the samples of said impulse response referred to said Reference HULA in the considered delay bin, and said spectrum is referred to said vertically folded impulse response; said computations are carried out through IFFT calculation with an oversampling factor

- said (Step 10.1.5 [STDAE-S4]) estimation of said horizontal spatial delay,

- (Step 10.1.6 [STDAE-S5]) checking whether combination of the impulse responses associated with said virtual antennas of said Reference HULA, after compensating for their phase differences on the basis of said estimated horizontal delay and of said estimated vertical delay, generates an amplitude peak in the absolute value of the resulting impulse response, named overall folded impulse response, wherein the impulse responses are associated with said virtual antennas of said sub-set of HULAs; and when the checking is positive execution of the following steps: said (Step 10.1.7) refinement of said range and complex amplitude of said most dominant point target; - said (Step 10.2 [STDAEC-S2]) cancellation of said most dominant point target,

- said (Step 10.3 [STDAEC-S3]) calculation of a residual energy in the considered delay bin and its comparison with a predetermined threshold {TSTDAEC)_t and checking whether said residual energy exceeds said predetermined threshold and, if so, return to said computation of the spectrum (Step 10.1.1) or to computation of the spectrum (Step 10.1.4), otherwise go to computation of the spatial coordinates (Step 11) otherwise, in negative case of said checking (Step 10.1.6) execution of the following steps:

- said (Step 10.4) updating of said acquired list of said delay bins and return to Step 10 for a further delay bin or exit when all delay bins are examined, wherein said execution (Step 10) ends when all the delay bins have been analyzed, then

- said (Step 11) computation of the spatial coordinates of all the point targets belonging to all the delay bins on the basis of said spatial delays and said complex amplitude and generation of an overall image of the propagation scenario.

20. Method according to claim 18 or 19, wherein said execution (Step 10 - [EMB#1]) is carried out by running multiple parallel processes whose number is equal to the number of said delay bins of said acquired list and said generation of the overall image (Step 11) is carried out when all parallel processes have been executed.

21. Method according to claim 18 or 19, wherein said acquisition (Step 8 - [EMB#2]) includes the acquisition of target energy and wherein said execution (Step 10 - [EMB#2]) is repeated sequentially over the acquired list, ordered according to said target energy, and wherein said generation of the overall image (Step 11) is carried out when all the delay bins have been analyzed.

22. Method according to claim 18 or 19, wherein said (Step 8) acquisition includes:

- (Step 8.1) acquisition of a list of a predetermined number of discretized ranges of said plurality of targets by selecting the indices of distinct delay bins where at least one point target is detected.

23. [STDAEC-S1] Method according to claim 18, wherein said estimation (Step 10.1) includes a preliminary step (Step 10.1.0) including: a) setting the iteration index i to 1, b) defining a matrix collecting the samples of the said impulse response available for each of said virtual antennas and corresponding to the considered delay bin only.

24. [STDAEC-S1] Method according to claim 19, wherein said estimation (Step 10.1) includes a preliminary step (Step 10.1.0) including: a) setting the iteration index i to 1, b) defining a matrix collecting the samples of the said impulse response available for each of said virtual antennas and corresponding to the considered delay bin only, c) extracting a vector, referring to said Reference VULA only, from said matrix.

25. [STDAE-S2] Method according to claim 19, wherein said vertical spatial delay estimation (Step 10.1.2) includes:

- (Step 10.1.2.1.0) representation of a target vertical spatial delay _v as sum of a main portion ( PV ^IDFT ) and a residual portion where F_IDFT where is the IFFT order, wherein said main portion is determined by searching the index ( b_n) of the vector element corresponding to the maximum absolute value of the elements of the vector and said residual delay computation is optional and can be found through an iterative procedure together with the corresponding target complex amplitude where a is said target amplitude and y is said target phase.

26. [STDAE-S4] Method according to claim 18 or 19, wherein said horizontal spatial delay estimation (Step 10.1.5) includes:

(Step 10.1.5.1.0) representation of a target horizontal spatial delay as sum of a main portion and a residual portion , where where is the IFFT order, wherein said main portion is determined by searching the index ( b_H ) of the vector element corresponding to the maximum absolute value of the elements of the vector and said residual delay computation is optional and can be found through an iterative procedure together with the corresponding target complex amplitude , where a is said target amplitude and y is said target phase.

27. [STDAE-S5] Method according to claim 18, wherein said estimate of the horizontal spatial delay (T_H) is employed to compute the phase difference between said reference antenna and said virtual antennas of said Reference HULA to be used in said constructive combination (Step 10.1.6) to compute said overall folded impulse response and its first three derivatives by combining the impulse response contributions referred to said sub-set of virtual antennas.

28. [STDAE-S5] Method according to claim 19, wherein said estimate of the vertical spatial delay and said estimate of the horizontal spatial delay are employed to compute the phase difference between said reference antenna and said virtual antennas of said sub-set of HULAs to be used in said constructive combination (Step 10.1.6) to compute said overall folded impulse response and its first three derivatives by combining the impulse response contributions referred to said sub-set of virtual antennas.

29. Method according to claim 18 or 19 wherein said refinement (Step 10.1.7) includes:

- (Step 10.1.7.1.0) representation of a target normalized delay where is the frequency step size of the employed radar, as sum of a main portion and a residual portion where where is the IFFT order, wherein said main portion is determined by searching the index of the vector element corresponding to the maximum absolute value of the elements of the vector and said residual delay is found through an iterative procedure together with the corresponding target complex amplitude where a is said target amplitude and is said target phase.

30. [STDAEC-S2-EMB#1] Method according to claim 18 or 19 wherein said cancellation (Step 10.2) includes the computation and subtraction of the contribution given by said dominant point target to the impulse response contribution referred to said considered delay bin to generate a new residual of the impulse response and of the corresponding derivatives.

31. [STDAEC-S2-EMB#2] Method according to claim 18 or 19 wherein said cancellation (Step 10.2) includes the computation and subtraction of the contribution given by said dominant point target to said impulse response and to the corresponding derivatives to generate a new residual of the impulse response and of the corresponding derivatives.

32. Method according to claim 18, wherein said computation (Step 11) includes the computation of the range and the azimuth (Q) as: and respectively, here, is the frequency step size of the employed radar, c is the speed of light, is the radar wavelength, is the distance between two adjacent virtual antennas of said HULA, is said discretized range, and is said horizontal spatial delay.

33. Method according to claim 19, wherein said computation (Step 11) includes the computation of the range the elevation and the azimuth as: and respectively, here, is the frequency step size of the employed radar, c is the speed of light, is the radar wavelength, is the distance between two adjacent virtual antennas of said VULAs and d_VH is the distance between two adjacent virtual antennas of said HULAs, is said discretized range, is said vertical spatial delay, and is said horizontal spatial delay.

34. [RPE] Method according any one of the previous claims from 18 - 33, further comprising the following preliminary steps:

- (Step 1: IFFT processing) estimation of said impulse response characterizing the communication channel associated with the said equivalent virtual antenna and its first derivatives with for each of said virtual antennas through IFFT calculation, with an oversampling factor (M),

- (Step 2: RVA selection) acquisition of a sub-set of a number (N_a) of virtual antennas impulse responses and of the corresponding derivatives,

Recursive execution of

(Step 3: STDREC-S1) calculation of an estimate of the parameters of the most dominant point target, including phase (y), amplitude (a) and delay (t) for each of said virtual antennas of said sub-set on the basis of said impulse response, and its derivatives, through IFFTs for each of said virtual antennas of said sub-set,

- (Step 4: STDREC-S2) cancellation of said most dominant point target and computation of a residual impulse response,

- (Step 5: STDREC-S3) computation of an energy of said residual impulse response and return to (Step 3) until said energy is over a predetermined threshold (T_STDREC)_r otherwise

(Step 6) computation of said average energy and fusion of ranges information of said point targets detected by said number of virtual antennas to list said point targets according to said average energy in a decreasing order.

35. Method according to claim 5 or 22, further comprising the acquisition of targets energy, wherein said acquisition includes:

- (Step 8.2) acquisition of a list of a predetermined number of targets energy corresponding to said discretized ranges.

36. Method for estimating the angular coordinates of mobile terminals served by a 5G base station, equipped with a plurality of receive (RX) antennas, arranged to receive downconverted radio frequency signals radiated by a single mobile terminal served by said base station, and radiating an orthogonal frequency division multiplexing (OFDM) signal that conveys pilot symbols; said method includes:

A pre-elaboration of said downconverted radio frequency signals including: a) acquisition of multiple signal samples of said downconverted radio frequency signals radiated by said single mobile terminal and computation of their FFT for each of said receive antenna, b) compensation for the phase and amplitude variations due to said pilot symbols transmitted at each frequency, for each of said receive antenna,

(Step 7) computation of the impulse response characterizing the communication channel associated with each of said receive antennas and its first derivatives for each of said receive antennas,

(Step 8) acquisition of a list of a predetermined number of discretized ranges, named delay bins, of said mobile terminals, (Step 9) selection of a reference antenna, a reference horizontal uniform linear array, named Reference HULA,

- (Step 10 [STDAEC]) sequential execution for each of said delay bins; said execution includes:

- (Step 10.1 [STDAEC-S1]) estimation of the horizontal spatial delay and complex amplitude of a mobile terminal, said estimation includes:

- (Step 10.1.4 [STDAE-S4]) computation of the spectrum and its first three derivatives of the sequence collecting the samples of said impulse response referred to said Reference HULA in the considered delay bin; said computations are carried out through IFFT calculation with an oversampling factor

- (Step 10.1.5 [STDAE-S4]) estimation of said horizontal spatial delay,

- (Step 10.1.6 [STDAE-S5]) checking whether combination of the impulse responses associated with said receiving antennas of said Reference HULA, after compensating for their phase differences on the basis of said estimated horizontal delay, generates an amplitude peak in the absolute value of the resulting impulse response, named overall folded impulse response; and when the checking is positive execution of the following steps:

- (Step 10.1.7) refinement of said range and complex amplitude of a single echo of the signal originating from said mobile terminal;

- (Step 10.2 [STDAEC-S2]) cancellation of said mobile terminal echo, (Step 10.3 [STDAEC-S3]) calculation of a residual energy in the considered delay bin and its comparison with a predetermined threshold (T_STDAEC), and checking whether said residual energy exceeds said predetermined threshold and, if so, return to computation of the spectrum (Step 10.1.4), otherwise go to computation of the spatial coordinates (Step 11) otherwise, in negative case of said checking (Step 10.1.6) execution of the following steps:

- (Step 10.4) updating of said acquired list of said delay bins and return to Step 10 for a further delay bin or exit when all delay bins are examined, wherein said execution (Step 10) ends when all the delay bins have been analyzed, then

(Step 11) computation of the spatial coordinates of all the mobile terminal echoes on the basis of said spatial delays and said complex amplitude and generation of an overall 2D map of possible positions of said mobile terminal.

37. Method according to claim 36, wherein said receive antennas form an array composed by one or multiple horizontal uniform linear arrays (HULAs) and one or multiple vertical uniform linear arrays (VULAs); the method including the following steps:

A pre-elaboration of said downconverted radio frequency signals including: a) acquisition of multiple signal samples of said downconverted radio frequency signals radiated by said single mobile terminal and computation of their FFT for each of said receive antenna, b) compensation for the phase and amplitude variations due to said pilot symbols transmitted at each frequency, for each of said receive antenna, said (Step 7) computation of the impulse response characterizing the communication channel associated with each of the said receive antennas and its first derivatives with for each of said receive antennas,

- said (Step 8) acquisition of a list of a predetermined number of discretized ranges, named delay bins, of said detected mobile terminals,

- said (Step 9) selection of a reference antenna, a reference vertical uniform linear array, named Reference VULA, a reference horizontal uniform linear array, named Reference HULA, and at least a sub-set with a number of said horizontal uniform linear arrays (HULAs) different from the Reference HULA,

- said (Step 10 [STDAEC]) sequential execution for each of said delay bins; said execution includes: said (Step 10.1 [STDAEC-S1]) estimation of the horizontal spatial delay and complex amplitude of a single echo originating from said mobile terminal and estimation of the vertical spatial delay; said estimation includes:

- (Step 10.1.1 [STDAE-SI]) computation of the spectrum of the sequence collecting the samples of said impulse response referred to said Reference VULA in the considered delay bin and its first 3 derivatives through IFFT calculation with an oversampling factor - (Step 10.1.2 [STDAE-S2]) estimation of said vertical spatial delay, - (Step 10.1.3 [STDAE-S3]) combination of the spectra associated with the Reference HULA and said sub-set of HULAs, after compensating for the phase differences between the Reference HULA and said sub-set of HULAs employing said estimate of the vertical spatial delay (^_^); said superposition, named vertical folding, generates a vertically folded impulse response, - said (Step 10.1.4 [STDAE-S4]) computation of the spectrum (^_^^^^,^) and its first three derivatives ({^_^^^^,^; ^ = 1, 2, 3}) of the sequence collecting the samples of said impulse response (^_^) referred to said Reference HULA in the considered delay bin, and said spectrum is referred to said vertically folded impulse response; said computations are carried out through IFFT calculation with an oversampling factor (^^{^}), - said (Step 10.1.5 [STDAE-S4]) estimation of said horizontal spatial delay, - (Step 10.1.6 [STDAE-S5]) checking whether combination of the impulse responses associated with said receiving antennas of said Reference HULA, after compensating for their phase differences on the basis of said estimated horizontal delay and of said estimated vertical delay, generates an amplitude peak in the absolute value of the resulting impulse response, named overall folded impulse response, wherein the impulse responses are associated with said receiving antennas of said sub-set of HULAs; and when the checking is positive execution of the following steps: - said (Step 10.1.7) refinement of said range and complex amplitude of a single echo of the signal originating from said mobile terminal; - said (Step 10.2 [STDAEC-S2]) cancellation of said mobile terminal echo, - said (Step 10.3 [STDAEC-S3]) calculation of a residual energy in the considered delay bin and its comparison with a predetermined threshold (^_^^^^^^), and checking whether said residual energy exceeds said predetermined threshold and, if so, return to said computation of the spectrum (Step 10.1.1) or to computation of the spectrum (Step 10.1.4), otherwise go to computation of the spatial coordinates (Step 11) otherwise, in negative case of said checking (Step 10.1.6) execution of the following steps: - said (Step 10.4) updating of said acquired list of said delay bins and return to Step 10 for a further delay bin or exit when all delay bins are examined, wherein said execution (Step 10) ends when all the delay bins have been analyzed, then - said (Step 11) computation of the spatial coordinates of all the mobile terminal echoes on the basis of said spatial delays and said complex amplitude and generation of an overall 3D map of possible positions of said mobile terminal.

38. Method according to claim 1 or 18, wherein in said selection (Step 9), said Reference HULA consists of a sub-set of adjacent and horizontally aligned virtual antennas.

39. Method according to claim 36, wherein in said selection (Step 9), said Reference HULA consists of a sub-set of adjacent and horizontally aligned receiving antennas.

40. Method according to claim 2 or 19, wherein in said selection (Step 9):

- said Reference VULA consists of a sub-set of adjacent and vertically aligned virtual antennas, - said Reference HULA consists of a sub-set of adjacent and horizontally aligned virtual antennas, each HULA of said sub-set of HULAs, different from the reference HULA, has the same size of the reference HULA.

41. Method according to claim 37, wherein in said selection (Step

9):

- said Reference VULA consists of a sub-set of adjacent and vertically aligned receive antennas,

- said Reference HULA consists of a sub-set of adjacent and horizontally aligned receive antennas, each HULA of said sub-set of HULAs, different from the reference HULA, has the same size of the reference HULA.

42. Method according to claim 1 - 2, or 18 - 19, further comprising the compensation for the amplitude distortion due to the non-uniform response of the antenna array for each of said virtual antennas on the basis of said refined complex amplitude (Step 10.1.7), and said estimated spatial frequencies, wherein said compensation includes:

- (Step 10.1.8) exploitation a deep neural network designed to solve a regression problem through a supervised learning approach wherein network training is based on a dataset collecting the amplitude of a number of said point targets; the compensation of the complex amplitude is carried out by multiplying said refined estimate of the complex amplitude (Step 10.1.7) by said estimated amplitude distortion for each of said virtual antennas.

43. A MIMO FMCW radar system including:

- a MIMO FMCW radar equipped with a plurality of transmitting (TX) and receiving (RX) antennas, arranged to generate real or complex signals in response to a propagation scenario including a plurality of point targets, the MIMO FMCW radar being arranged to generate real or complex signals in response to a propagation scenario including a plurality of point targets and - computing means arranged to receive signals generated by said transmitting and receiving antennas and model each couple of said transmitting and receiving antennas with an equivalent model defining virtual antennas composed by one or multiple horizontal uniform linear arrays (HULAs) and, optionally, to one or multiple vertical uniform linear arrays (VULAs) and to perform all the steps of any one of the previous claims from 1 to 17 and 35, 38, 40, 42.

44. A MIMO SFCW radar system including:

- a MIMO SFCW radar equipped with a plurality of transmitting (TX) and receiving (RX) antennas, arranged to generate complex signals in response to a propagation scenario including a plurality of point targets and

- computing means arranged to receive signals generated by said transmitting and receiving antennas and to associate to each couple of said transmitting (TX) and receiving (RX) antennas to an equivalent model defining virtual antennas; said virtual antennas forming a virtual array composed by one or multiple horizontal uniform linear arrays (HULAs) and, optionally, one or multiple vertical uniform linear arrays (VULAs), and to perform all the steps of any one of the previous claims from 18 to 35, 38, 40, 42.

45. A 5G base station system including

- a plurality of receive (RX) antennas, arranged to receive downconverted radio frequency signals radiated by a single mobile terminal served by said base station, and radiating an orthogonal frequency division multiplexing (OFDM) signal that conveys pilot symbols, and computing means arranged to receive downconverted radio frequency signals operatively connected with said plurality of receive (RX) antennas and configured to execute all the steps of claims 36 or 37.