3

Waveforms, Matched Filtering,

and Radar Signal Processing

3.1 INTRODUCTION
This chapter describes the waveforms employed by radars, introduces the con-
cept of the optimal matched filter processor and discusses some common signal
processing implementations used in phased-array radars. Topics covered in-
clude:

• Representations of complex-valued waveforms

• Fourier transforms of radar signals and their properties
• The matched filter:
– Simple pulsed continuous wave (CW)
– Linear frequency modulation (LFM)
– Phase-code modulation
– Ambiguity function and diagrams
• Signal processing implementations:
– All-range digital pulse compression
– Spectrum analysis or stretch processing of LFM

55
56 Chapter 3

– Phase code processing

– Constant fals alarm rate (CFAR) detection processing
– Monopulse processing.

There are many excellent references for these topics in the literature. These in-
clude references [1], [2], [10], and [12] for general background, while references
[5–9], [11], [13], and [15] specifically cover signal theory and signal processing.

3.2 COMPLEX-WAVEFORM REPRESENTATIONS

A waveform can be represented mathematically as:

 ( t ) = r ( t ) cos  2π f0 t + θ ( t ) , (3.1)

where r(t), f0, and q(t) are the envelope modulation, operating or center fre-
quency, and the phase modulation, respectively [6, 8]. Using the exponential rep-
resentation for the cosine yields:

{ (
 ( t ) = r ( t ) exp j  2π f0 t + θ ( t ) ) (
+ exp − j  2 π f0 t + θ ( t ) )} / 2 ,
(3.2)

where l(t), like in the first representation, is still “real valued.”

A complex representation of can be stated as:

{
ν ( t ) = r ( t ) exp j  2π fo + θ ( t )  . } (3.3)

Note that:

s ( t ) = ν ( t ) exp  − j ( 2π fo t ) = r ( t ) exp  jθ ( t ) . (3.4)

s(t) is termed the “complex envelope” of the original waveform, l(t). s(t) is an ap-
proximation to l(t) for narrowband signals but is close enough for most radar
analysis.
 Waveforms, Matched Filtering, and Radar Signal Processing 57

The true complex waveform envelope is defined using Hilbert transforms.

Also note that:

 (t ) =
1
2
{ {
Re r ( t ) exp j  2π fo t + θ ( t )  }} . (3.5)

The complex notation in equation (3.3) through (3.5) and the waveform’s com-
plex envelope are useful when analyzing signal processing effects; in other
words:

• Use the complex envelope, s(t), for calculations

• Convert back to real-valued waveforms, such as l(t), after otherwise “messy”
mathematics are completed.

3.3 FOURIER TRANSFORMS

Fourier transforms and transform theory are useful for manipulating wave-
forms and signal processing. The Fourier transform is defined for finite energy
2
signals (i.e., for signal x(t), ∫ x ( t ) dx must be finite) [8]. Note that this is differ-
ent from the Fourier series, where signals are of infinite energy but are of finite
power.
The Fourier transform is defined as:

∞
X( f ) = ∫ x ( t ) exp ( − j 2π f t ) dt ,
∞
(3.6)

with the inverse Fourier transform given by:

∞
x (t ) = ∫ x ( f ) exp ( j 2π f t ) df .
∞
(3.7)

For example, if x(t) is defined as:

x (t ) = { 1 ,
0,
t
t
≤ τ 2
> τ 2
(3.8)
58 Chapter 3

then its Fourier transform is given by:

X ( f ) = ∫ τ 2 exp ( − j 2π f t ) dt = τ sin ( f πτ ) / ( f πτ ) ≡ sin c ( f τ ) .

τ 2
(3.9)

Some important and useful Fourier transform properties for time or frequency
shifting are:

x (t ) ⇔ X ( f ) where ⇔denotes Fourier Transform

x ( t − tο ) ⇔ exp ( − j 2π f tο ) X ( f ) . (3.10)

X ( f − fο ) ⇔ exp ( j 2π fο t ) x ( t ) .

The Fourier transform properties associated with auto correlation (or cross-
correlation) are:

∞
Rx (τ ) = ∫ x ( t ) x * ( t − τ ) dt ,
∞
or: ∞
. (3.11)
Rx (τ ) = ∫ X ( f ) exp  j 2π f τ  df
2

Similarly, the transform properties of the often-used Dirac delta function, δ(t),
are given by:

∞ ∞
f (0) = ∫ ∞
f ( t ) δ ( t ) dt ⇒ Ι( f ) = ∫ δ ( t ) exp  − j 2π t  dt
∞
∞ ∞
. (3.12)
f ( t0 ) = ∫ f ( t ) δ ( t − tο ) dt ⇒ δ ( t ) = ∫ ∞exp  j 2π f t  df
∞

3.4 MATCHING FILTERING

Consider the output of a filter with impulse response h(t) and input s(t):

∞
 0 (t ) = ∫ s ( t ) h (τ − t ) dt .
∞
(3.13)
 Waveforms, Matched Filtering, and Radar Signal Processing 59

Using the inverse Fourier transform, equation (3.13) can be expressed as:

∞
 0 (t ) = ∫ H ( f ) S ( f ) exp ( j 2π f t ) df .
∞
(3.14)

The optimum filter with respect to maximizing the output signal-to-noise ratio
(SNR) in white noise is termed the “matched filter” to s(t), and denoted hMF(t) or
in the frequency domain representation as HMF (f). The matched filter impulse re-
sponse is given by [6]:

hMF ( t ) = g∗ s ∗ ( − t ) ⇔ H MF ( f ) = g S ∗ ( f ) , (3.15)

where g is a complex constant.

The output of the matched filter to an input of s(t) is given by:

∞ ∞ ∞
 MF ( t ) = ∫ h( x ) s (t − x ) dx
∞
= ∫ ∞
h (t − y) s ( y) dy = ∫ s * (t ′ − t ) s(t ′) dt ′ .
∞

(3.16)

∞
or:  MF ( t ) = ∫ ∞
s (t ) s *(t − τ ) dt . (3.17)

Equation (3.17) is the cross-correlation of s(t) and s*(t). If the output noise power,
Pnoise, is defined as:

Pnoise = ( 2) ∫
No ∞

∞
H MF ( f ) df ,
2
(3.18)

then the SNR at the output of the matched filter is given by:

SNRMAX = ∫
∞

∞
S ( f ) df /
2
(N 2 )
o . (3.19)

Since the output of the matched filter is:

∞
 MF ( t ) = ∫ s (t ) s *(t − τ ) dt ,
∞
(3.20)
60 Chapter 3

it can be represented equivalently as:

∞
 MF ( t ) = ∫ S ( f ) S ( f ) exp ( j 2π f t ) df .
∞
∗
(3.21)

A conceptual block diagram of equation (3.21) is shown in Figure 3.1. This is a

frequency-domain implementation of the matched filter and is equivalent to a
time-domain correlator.
Consider a rectangular pulse:

t
s ( t ) = rect   ⇒ .
τ

The matched filter output from equation (3.20) is given by:

∞ ∞
t t −τ
 MF ( t ) = ∫ s ( t ) s * ( t − τ ) dt
∞
= ∫ rect  τ  rect
∞
*
  dt
τ 
   t    t − τ  
= F −1  F rect    F rect 
τ    τ    {
= F −1 τ 2 sin c 2 ( f τ )  . }
  
 τ 
= τ 2 triang   ⇒
 2τ 

(3.22)

Figure 3.1 Conceptual Block Diagram of Matched Filter Processor

 Waveforms, Matched Filtering, and Radar Signal Processing 61

If the signal has voltage spectrum:

F
E0 ( f ) = sin (π fB ) (π fB ) ⇔ rect [ Bt ] ∗ rect [ Bt ]
2

 o ( t ) = triang ( Bt ) ⇒

(3.23)

A unity-magnitude linear frequency modulation pulse has a complex enve-

lope given by:

sLFM ( t ) = )
exp ( jπ k t 2 ; 0 ≤ t ≤ τ , (3.24)

where the instantaneous frequency is defined as:

 1  dφ
f (t ) =   = kt , (3.25)
 2π  dt

and where its maximum is given by:

∆ = kτ . (3.26)

The LFM matched filter output can be calculated as:

t
 MF ( t ) = triang   sin π ∆t (1 − t τ )
π ∆t (1 − t τ )
, (3.27)
τ
 

which has an approximate “compressed” pulse width of 1/∆ and where near the
main lobe of the output the time response is approximately sin x/x. Figure 3.2 il-
lustrates the frequency spectrum of an LFM waveform. The complex-conjugate
of this spectrum is effectively the desired matched filter frequency response.
The LFM waveform spectrum is given by:

∞
S( f ) = ∫ ∞
exp( jπ k t 2 ) exp( − jπ f t ) dt , (3.28)
62 Chapter 3

Figure 3.2 LFM Frequency Spectrum (Courtesy of D. P. Harty)

where:

 f
S( f ) ≈ rect   , (3.29)
 ∆

with phase:

φ( f ) ≈ (2π f )2
4π k , (3.30)

so that:

 f
S ( f ) ≈ rect   exp  jφ ( f ) . (3.31)
 ∆
 Waveforms, Matched Filtering, and Radar Signal Processing 63

Hence, the matched filter to a LFM waveform for large time-bandwidth prod-
ucts is approximately given by:

 f
H ( f ) = rect   exp  − jφ ( f ) (3.32)
 ∆

and:

 f F sin (π ∆t )
EOUT ( f ) ≈ rect   ⇔  OUT ( t ) ≈ c . (3.33)
 ∆ π ∆t

3.5 WAVEFORM AMBIGUITY DIAGRAM

The ambiguity diagram is a three-dimensional plot that shows the results of con-
volving frequency frequency-shifted signals with a fixed reference signal, that is,
with the matched filter impulse response. The Doppler frequency-shifted signals
represent waveform echoes from moving objects. ([6, 12, 16] for more detailed
coverage of this topic.)
An ambiguity diagram depicts two important properties of a waveform: infor-
mation about a waveform’s inherent capability to resolve targets in range and
Doppler. The shape of the ambiguity diagram indicates how well suited a wave-
form is to determine the range and range-rate of an object. The horizontal axes of
the diagram are labeled as range (or time delay) and Doppler frequency, and the
vertical axis represents magnitude in decibels (dB).
The ambiguity diagram for an LFM pulse is shown in Figure 3.3. The figure il-
lustrates the inherent range-Doppler coupling of the LFM waveform.

3.6 THE FAST FOURIER TRANSFORM (FFT)

The fast Fourier transform (FFT) is an efficient implementation of the discrete
Fourier transform commonly used for sampled data. References [7–9] are excel-
lent sources for this material. This section introduces the concept. The FFT is de-
fined as:
64 Chapter 3

Figure 3.3 Ambiguity Diagram for a LFM Waveform (Courtesy of D. P. Harty)

N −1
Α( k ) = ∑α
n=0
n e− 2π j n k / N . (3.34)

The “Cooley-Tukey” FFT algorithm first rearranges the input elements in bit-
reversed order, then builds the output transform (decimation in time). Efficiency
is achieved by breaking a transform of length N into two transforms of length
N/2 using the identity where i is used in place of the j used in equation (3.34):

. (3.35)
 Waveforms, Matched Filtering, and Radar Signal Processing 65

3.7 DIGITAL IMPLEMENTATION OF MATCHED FILTERS

Figure 3.1 can be implemented for sampled-data in discrete-time using FFTs
[7–9]. This is illustrated in Figure 3.4.
Note that Figure 3.4 is a digital implementation of an all-range LFM pulse
compression approach. When bandwidths are too large to allow sampling at or
above the Nyquist rate, this approach cannot be employed. A technique referred
to as “stretch” processing of “spectrum analysis” can be used in these cases, ref-
erence [14]. This is depicted in Figure 3.5.
There is a tradeoff between achievable analog-to-digital converter (A/D) sam-
pling rates (and number of bits) and the size of the range window (RW in previ-
ous diagrams) to be processed, where the RW is the part of range space over
which pulses can be compressed or match filtered.
For a very high LFM bandwidth requirement the pulse cannot be sampled fast
enough since A/Ds are not available that run at the required Nyquist rate or
higher. Therefore, for very wideband waveforms, all-range digital pulse com-
pression (i.e., two-pass FFT processing) cannot be used. This is the reason for use
of spectrum analysis-type techniques (or “stretch” processing), including its
one-pass FFT processing implementation where lower A/D rates can be used
but limited to smaller RWs (see reference [14]).
When the LFM bandwidth is within the range of current achievable A/D pa-
rameters (sampling rates, numbers of bits), then all-range processing (i.e., large
RWs) is feasible and is usually used. However, larger RWs do require larger FFT
sizes and the associated throughput requirements. For these reasons, the two-
pass FFT approach is usually used only for:

• Search (e.g., LFM bandwidths ~ 1 MHz or less) when large range extents are
required
• Tracking (e.g., LFM bandwidths ~ 5 to 20 MHz) when large extents are neces-
sary.

When the tracking bandwidth is only 10 or 20 MHz, but only a small RW is

necessary (like reacquisition or other functions where the range uncertainty is
relatively small), then “stretch” might be used even for lower-bandwidth wave-
forms. Some radars will employ two types of track processing:
66

Figure 3.4 LFM Matched Filter Implementation Using FFTs

 Figure 3.5 Stretch or Spectrum Analysis for LFM Matched Filtering
67
68 Chapter 3

• A large RW type (larger FFTs required) that uses two-pass (all-range process-
ing)
• A small RW type that uses “stretch” (smaller FFTs required)—this allows
many more tracks to be processed.

3.8 PHASE-CODED WAVEFORMS

The LFM waveform achieves a wide bandwidth using analog modulation (or
continuous modulation). Phase-coded waveforms obtain a wide bandwidth by
using a discrete-phase coding approach. A general phase-coded waveform is of
the form shown in Figure 3.6.
The phase, φi, can be selected from a discrete set of length N, where might be N
= 2m where m = 1, 2,..., M. If N = 2, the waveform is bi-phase, if N = 4, it is quad-
phase, etc.
For bi-phase coding:

 3π 
φi = { 0,π }; for quad − phase,φ i =  0 , π 2, , π ,

,
2 
(3.36)

The same basic matched filter definition is used for phase-coded waveforms,
only in discrete form. Hence, integration becomes summation, and the convolu-
tion integration is replaced by discrete correlation function:

n
 OUT ( k ) = ∑ s ( k ) s (n − k ) .
*
(3.37)
k =1

Figure 3.6 Depiction of a Phase-Coded Waveform

 Waveforms, Matched Filtering, and Radar Signal Processing 69

The correlator in equation (3.37) can be implemented using a tapped delay

line (TDL), where the tap weights are the time-reversed phase modulations as
shown in Figure 3.7.
The equivalent bandwidth, Be, of the phase-coded waveform is defined as:

1
Be ≈ . (3.38)
τs

3.9 WAVEFORM SCHEDULING

Figure 3.8 illustrates some fundamental waveform scheduling concepts, in par-
ticular some key definitions. Specifically, the concepts of minimum and maxi-
mum range, duty factor, and receive window are defined.
Figure 3.9 illustrates the important concept of timeline occupancy. Occupancy
can be considered a radar resource, just as duty factor. In fact, these two quanti-
ties are usually those that are “managed” by a radar resource manager and
scheduler function.
The example at the bottom of Figure 3.9 represents a radar PRF of 25 pulses/
second or, alternatively, 25 beams/second. A radar is defined to be “occupancy-
limited” for specified search if:


Re quired

ψ 
Available

> PRF , (3.39)

Ω Tsc

Figure 3.7 Matched Filter for a Phase-Coded Waveform

70 Chapter 3

Figure 3.8 Waveform Scheduling Concepts and Definitions

Figure 3.9 Illustration Radar Timeline Occupancy Concept

where Ψ, Ω, and TSC are the search volume in radians2, the radar antenna beam
area in radians2, and the specified scan or search frame time, respectively, and
PRF is the pulse repetition frequency.

3.10 WAVEFORMS AND RADAR FUNCTIONS

The waveforms that are typically employed for common radar functions are:
 Waveforms, Matched Filtering, and Radar Signal Processing 71

• Search:
– Un-coded CW
– Narrow band linear frequency modulation (LFM)
– 500 kHz to 1 MHz bandwidth (i.e., very narrowband)

• Track initiation/track maintenance:

– LFM
– 5 to 20+ MHz bandwidth (i.e., narrowband)

• Target classification:

– LFM
– Wide bandwidths.

As can be seen, very narrowband waveforms are usually employed for search,
narrowband waveforms are used for tracking, and wideband waveforms are
used for classification.

3.11 OTHER RADAR SIGNAL PROCESSING FUNCTIONS

This section describes two signal processing functions in addition to the wave-
form matched filtering described in previous sections. Constant false alarm rate
(CFAR) and monopulse processing are described in the following subsections.

3.11.1 Constant False Alarm Rate Processing

CFAR processing is a signal processor algorithm that sets the detection
threshold for declaring target returns [6, 17]. This selection of the detection
threshold affects the false alarm performance and detection performance of a
radar system. In thermal noise-only environments (no clutter or jamming),
the detection threshold can be calculated from equations (3.40) and (3.41),
where PFA and 2σ2 are the probability of false alarm and the thermal noise
power, respectively:

α VT 2
∞ −   22 −

∫
σ 
PFA = α
σ2
e dα = e 2σ 2
(3.40)
− VT
72 Chapter 3

and:

VT 2 = − 2 σˆ 2 n ( PFA ) . (3.41)

Figure 3.10 shows the Rayleigh probability density associated with the random
thermal noise magnitudes.
However, in most applications, the noise power neither is known precisely nor
is constant. This necessitates estimating the localized (in range and angle) noise
floor before computing the threshold using measurements of the background
noise.
When the environment consists of noise plus homogeneous or nonhomoge-
neous clutter and/or jamming, the estimates of background interference are
needed. The CFAR processor is designed to perform these functions—that is, to
estimate the background interference level and “bias” it to achieve the desired PFA.
Figure 3.11 depicts an example of a cell-averaging CFAR, referred to as a CA-
CFAR. A general configuration for a CFAR is depicted in Figure 3.12. Received
target echoes are amplified (linearly or logarithmically), detected, and then
passed through a tapped delay line. Usually “early” and “late” background
noise or interference levels are estimated using the “late” and “early” TDL cells
used with some combining logic (usually selectable). The “cell under test”
(CUT) is compared with the calculated threshold using a “biased” noise/inter-
ference estimate to achieve the desired PFA.

Figure 3.10 Probability Density for Thermal Noise

 Figure 3.11 Example of a Cell-Averaging CFAR (CA-CFAR)
73
74 Chapter 3

Figure 3.12 Generalized CFAR Processor Block Diagram

The most common form of CFAR is the cell-averaging type (i.e., CA-CFAR).
Usually “early” and “late” cells are used to compute early and late “window”
noise averages. These may be simple arithmetic averages, or they may seek to
exclude some cells from the computed average. Typically excluded cells include:

• Cells on either side of the CUT (e.g., 1 to 3 or so)

• Cells that exhibit large values (spiky interference or noise); often the largest
two or three values are excluded.

If M cells remain after “editing,” the noise estimate (for linear) processing is
given by:

Pr =  Σ qm  M .
M
(3.42)
 m =1 

Note that for logarithmic processing, the estimate may be computed using:

P r − log = Σ M
log q  M . (3.43)
 m =1 2 m 
 Waveforms, Matched Filtering, and Radar Signal Processing 75

The logarithm of any base can be used in place of the base-2 logarithm indicated
in equation (3.43).
Since the noise (or interference) power estimate uses a finite (M) number of
cells, there is a loss in detectability, on average, when using the linear CFAR pro-
cessor for detection. For Swerling I type target fluctuations, the loss is reflected
in the PD:

P 
M n  FA 
   −
1
    Pd 
PD = (1 + γ a ) 1 + γ a +  P M
− 1   ; where γ a = .
n ( PD )
FA
     
(3.44)

This can be compared to the ideal PD in noise for Swerling I targets:

1 1
PD = ( PFA ) 1 + SNR = ( PFA ) 1 + γ a . (3.45)

The signal-to-interference plus noise ratio (SINR) for M-cell estimates is there-
fore:

 1
  1

γ a ( M ) = ( PD PFA ) M − 1  1 − PD
M
 . (3.46)
 

The difference between γa(M) and the ideal

n ( PFA )
SNR = −1
n ( PD )

is referred to as the (homogeneous environment) CFAR loss, which in decibels is

approximately given by:

L CFAR ( dB ) = 10 log10  γ a ( M ) γ a ( ∞) . (3.47)

76 Chapter 3

In the previous homogeneous environments, it was assumed that the back-

ground was statistically “flat” or having a near constant mean with a rela-
tively small variance. However, the interference background can often consist
of discrete values (such as other targets), extended clutter, and other discrete
returns. In these environments, the false alarm probabilities achieved will
vary, sometimes considerably from their desired value. In addition, the PD
may not increase (i.e., reduce the CFAR loss) when larger numbers of cells are
used.
Alternative techniques are required to achieve the desired CFAR false alarm
performance. However, this performance is obtained at the cost of a greater loss
in detectability. Examples of nonhomogeneous backgrounds include: step func-
tions, ramps (up-constant-down) or trapezoidal envelopes.
The CA-CFAR cannot maintain constant false alarm performance during sig-
nificant changes in background levels (e.g., transitions for step functions, ramps
up and down for trapezoidal shapes). The PD obtained in nonhomogeneous
backgrounds, such as “clumps” of clutter, is given by:

−1
   −1  
 P0 (1 + γ a )  ,
M
PD = π  1 + Pm  PFAM − 1    (3.48)
m =1
    

where:
P0 = total interference power of test cell
Pm = total interference of reference cell M

Note that if test and reference cells all have the same powers, this reduces to
the previous result:

M
   − M1   
PD = (1 + γ a ) 1 + γ a +  Pfa − 1   . (3.49)
     

The linear CFAR, although reasonably optimum for homogeneous back-

grounds is suboptimum for nonhomogeneous backgrounds, due to sharp transi-
tions in background interference levels. This suboptimum estimation
 Waveforms, Matched Filtering, and Radar Signal Processing 77

performance can be improved by using the logarithmic CFAR. For this type of
CA-CFAR, with “ramp” or transition portions of the interference background,
the false alarm probability is independent of test cell power and slope of the
ramp.
The required SINR necessary to achieve a given PD and PFA for logarithmic
CFAR is greater than that for the linear CA-CFAR in homogeneous background.
Hence, geometric mean (or logarithmic CA-CFAR) processors exhibit higher
CFAR losses. However, the higher loss is traded off against ensuring the desired
PFA performance in nonhomogeneous interference backgrounds.
Although CA-CFAR is optimum in homogeneous noise and jamming, the op-
timality of geometric mean CFAR is not known. Also, the type of CFAR neces-
sary to achieve “best” performance varies with the type of background present.
Alternative CFARs developed to deal with different nonhomogeneous interfer-
ences include:

• Greatest-of (GO-CFAR)
• Order Statistics (OS-CFAR)
• Censoring (already discussed)
• Composite (combinations of the above and other forms)

“Greatest-of” uses the larger of the “early” and “late” averages to compute the
threshold. This results in a slight increase in CFAR loss when used in homoge-
neous environments.
Other types of nonhomogeneity are the presence of additional targets in the
reference cells. An M-cell CA-CFAR with K additional targets present results in a
degraded PD of:

−(M − K )
PD = 1 + ka (1 + γ A ) 1 + ka (1 + γ ) 1 + γ A 
−K

  , (3.50)

where ka is the adaptive threshold multiplier, γ is the SINR of the test cell (CUT),
and γA is the SNR of the additional targets. The multiple-target-in reference cells
condition results in higher CFAR losses. The loss can be reduced by using the
geometric mean, censoring (discussed previously), use of the median to form the
threshold, or more generally using a multiple of an arbitrary order statistic of the
78 Chapter 3

reference cells (OS-CFAR). The CFAR loss for a 32-cell CFAR for PFA = 10–6 is
given by:
CA-CFAR 0.97 dB
GO-CFAR 1.13 dB
OS-CFAR (75% rank) 1.45 dB (3.51)
GO-OS-CFAR (75% rank) 1.66 dB
CA-CFAR, 1-cell censor 1.01 dB
CA-CFAR, 2-cell censor 1.06 dB.
For slowly varying clutter, a smoothed (recursive) clutter map can be com-
puted for range and angle cells. The recursive estimate is, for the kth cell found
using:

Qk ( m ) = (1 − w ) Q ( m − 1)
k + w qm ( k ) , (3.52)

where qm(k) is the kth cell amplitude, w is the smoothing weight, and m is the
“scan” number. For this clutter map estimate, Pd is given by:

∞ −1
PD = π  1 + kc w (1 − w ) (1 + γ ) 
r
, (3.53)
r = 
a

where kc is the threshold multiplier for a given PFA (iteratively determined with
γa set to zero).
The objective of using composite techniques is to reduce the CFAR loss. Effec-
tively, by switching CFAR types based on measured averages compared to a
threshold (or other rule), the lowest-loss CFAR that achieves desired PFA perfor-
mance is employed. An example would be combining a CA-CFAR for noise-only
and a censoring CFAR for multiple targets present, or log CA-CFAR for
“clumps” of clutter.

3.11.2 Monopulse Processing

Another key technique used by radars is monopulse. Monopulse is used to ex-
tract angle measurements target returns [18]. The maximum-likelihood angle es-
timate is given by:
 Waveforms, Matched Filtering, and Radar Signal Processing 79

θ  ∆ 
θˆML = 3 Re   , (3.54)
km  Σ 

where θ3, km, Σ, and ∆ are the antenna 3 dB beamwidth, monopulse slope, and
antenna sum and difference-channel voltages, respectively, and Re{} denotes the
real part of a complex number.
Monopulse accomplishes this function by separating two-dimensional receive
antennas into azimuth and elevation “quadrants” and combines them to esti-
mate angular positions. An “error” pattern is formed as e(θ) = ∆(θ)/Σ(θ) as
shown in Figure 3.13.
The amplitude and sign of e(θ) indicates the distance of the target from the an-
tenna pointing angle.
The thermal noise-limited one-sigma angle accuracy associated with the
monopulse technique is the familiar Cramer-Rao-type bound:

Figure 3.13 Diagram of Basic Monopulse Technique

80 Chapter 3

θ3 1
σθ = , (3.55)
km 2 SNR

where SNR is the signal-noise-ratio at the matched filter and monopulse com-
parator output.
The complex-valued monopulse ratio also can provide an indication if there
are two or more unresolved targets in a range-cell but within the 3 dB antenna
receive beamwidth. A hypothesis test similar to that used for target detection
can be employed for this purpose:

H1
 ∆  >
γ imaginary = Im   T unresolved , (3.56)
 Σ  <
H0

where Im{} denotes the imaginary part of a complex-valued quantity, H0 and Hi

are the hypotheses for resolved and unresolved targets, respectively, and Tunre-
solved is the test threshold. In a similar manner to the target detection problem, the
threshold is selected to achieve a specified probability of false “unresolved” in-
dication. Target returns with values of γimaginary that exceed the threshold would
not be used for tracking or discrimination, for example, where the corrupted re-
turn data could potentially degrade radar performance.

3.12 REFERENCES
[1] D. K. Barton, Radar System Analysis and Modeling, Artech House, 2004
[2] E. Brookner, Aspects of Modern Radars, Artech House, 1988
[3] S. Haykin, Adaptive Radar Signal Processing, Wiley-Interscience, 2006
[4] S. Kay, Modern Spectral Estimation: Theory and Application, Prentice-Hall, 1999
[5] N. Levanon, Radar Principles, Wiley-Interscience, 1988
[6] R. Nitzberg, Radar Signal Processing and Adaptive Systems, 2nd Edition, Artech House,
1999
[7] A. Oppenheim & R. Shafer, Digital Signal Processing, Prentice-Hall, 1975
[8] A. Papoulis, Signal Analysis, McGraw-Hill, 1977
[9] L. Rabiner, Theory and Application of Digital Signal Processing, Prentice-Hall, 1975
[10] M. Skolnik, Introduction to Radar Systems, 3rd Edition, McGraw-Hill, 2002
[11] S. Haykin & A. Steinhardt, Adaptive Radar Detection and Estimation, Wiley, 1992
[12] D. K. Barton, Modern Radar System Analysis, Artech House, 1988
 Waveforms, Matched Filtering, and Radar Signal Processing 81

[13] D. R. Wehner, High Resolution Radar, Artech House, 1987

[14] W. J. Caputi, “Stretch: A Time Transformation Technique,” IEEE Transactions on
Aerospace Electronic Systems, AES-7, March 1971
[15] J. V. Candy, Signal Processing—The Modern Approach, McGraw-Hill, 1988
[16] A. W. Rihaczek, Principles of High-Resolution Radar, Mark Resources, 1977
[17] G. Minkler & J. Minkler, CFAR, Magellan, 1990
[18] S. M. Sherman, Monopulse Principles and Techniques, Artech House, 1984

3.13 PROBLEMS
1. Consider a radar waveform with a triangular-shaped pulse as shown below.
Calculate the matched filter output for this waveform. Sketch the matched
filter output waveform (hint: start with the analysis from chart 10 and apply
the convolution property of Fourier transforms).

2. Estimate the approximate 3 dB range resolution of the matched filter output

in problem 1 if τ = 10 µsec. What is the approximate bandwidth of the
matched filter output?

3. If a linear FM (or “chirp”) waveform is used to achieve the above range reso-
lution, what bandwidth is required? If no time sidelobe weighting is applied
during matched filtering, compare with the sidelobes of the waveform from
problem 1. Is the waveform from problem 1 realizable? If a triangular taper
is applied to the LFM waveform, what is the amplitude of the first sidelobe?
What penalties are incurred when using the triangular taper?

4. Consider a radar that employs a 1 millisecond (τ) 1 MHz LFM for search.
All-range digital pulse compression is to be used for matched filtering to
cover a 10 km range window (RW). Since some degree of interpolation is de-
sired, a complex sampling rate of 1.2 MHz is to be used. Calculate the size of
the fast Fourier transform (FFT) required for the signal processing (assuming
82 Chapter 3

that zero-filling will be used to pad out the FFT to a power of 2, i.e., 2M ≥ Ns).
For calculation purposes, note that for pulse compression:

Ns = ( f ) (τ
samp p + RW ) N FFT = ( 2) M .

5. For tracking purposes, a 1 millisecond 20 MHz LFM is used. A “stretch” or

“spectrum analysis” approach is employed for matched filtering over the 1
km range window. Calculate the required complex sampling rate necessary
to implement this technique and the required FFT size. For calculation pur-
poses, note that for the “stretch” process:

 B
Bs =   ( RW ) .
τp

6. Consider a CFAR processor that uses a 16 range cells to estimate the back-
ground noise-plus-interference level. For a required probability of false
alarm of 10–6, calculate the resulting probability of detection when probabil-
ity of detection using the ideal noise threshold is 0.9 for a Swerling I target.

P 
M n  FA 
   1
    PD 
(1 + γ a )
−
PD = 1 + γ a +  P
M
− 1   ; where γ a = .
n ( PD )
FA
    

7. Estimate the sensitivity loss (i.e., loss in equivalent SNR) that results when
using the 16-cell CFAR. Note the relationship for a Swerling I target using an
ideal noise threshold is:

1
PD = ( PFA ) 1 + SNR .
8. Consider the CFAR processor and parameters used in problem 1 in a non-
homogeneous environment. Assuming that the background level is identical
in each range cell, calculate the resulting probability of detection and com-
pare with the result in problem 6.
 Waveforms, Matched Filtering, and Radar Signal Processing 83

M
   −1   
PD = (1 + γ a ) 1 + γ a +  PFAM − 1   .
     

9. Consider the use of a 32-cell CFAR processor to detect a Swerling I target in a

homogeneous noise background. If the SNR is 15 dB and the desired proba-
bility of false alarm of 10–6, calculate the resulting probabilities of detection
when cell-averaging (CA) CFARs with and without 2-cell censoring, a great-
est-of (GO) CFAR, and a GO-ordered statistics (OS) CFAR are used. Com-
pare this to the case of an ideal noise threshold as calculated in problem 7.