JP2002328162A - Target tracking device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a target tracking device providing favorable tracking accuracy by controlling sizes and center distances of error ellipsoids and the number of models so the error ellipsoids constantly overlap, and suppressing unnecessary compensation to position and speed occurring when a turn duration of a target is shorter than a sampling interval. SOLUTION: The target tracking device is provided with an error ellipsoid multiplicity determining means of determining multiplicity of the error ellipsoids which are equal probability lines showing existence expectation probabilities of the target around predicted positions of N pieces of kinetic models on the basis of predicted values and a predicted error covariance matrix per kinetic model calculated by a Kalman filter, and an error ellipsoid enlarging/reducing means of enlarging and reducing the error ellipsoids so there are no blank areas between error ellipsoids on the basis of the multiplicity of the error ellipsoids determined by the error ellipsoid multiplicity determining means.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a target tracking device for estimating target movement parameters such as target position and speed based on observation data on target position information from sensors and the like. TECHNICAL FIELD The present invention relates to a target tracking device that keeps stable tracking accuracy for any target by mixing reliability of a plurality of motion models by considering the overlap of probability density distributions of motion models.

[0002]

2. Description of the Related Art FIG. 12 is a block diagram showing a conventional target tracking device disclosed in Japanese Patent Application Laid-Open No. Hei 1-292678. FIG. 13 is a flowchart for explaining the operation of the target tracking device shown in FIG. Hereinafter, the configuration of the conventional target tracking device shown in FIG. 12 will be described.

In FIG. 12, reference numeral 1 denotes observation information on a target position observed by a sensor (hereinafter referred to as “target position observation information”).
Is a target observation device that acquires a target tracking target based on the target position observation information, the predicted position, and the prediction error covariance matrix described above, and calculates the reliability of the obtained observation value. A gate judging unit for outputting to the unit 3 and the smoothing unit 4.

[0004] Further, 3 calculates the reliability of each motion model based on the above-mentioned observed value, the predicted value for each motion model, the prediction error covariance matrix for each motion model, and the reliability of each motion model one sample before. Is a reliability calculator that calculates
The reliability of each motion model before sampling is stored in the reliability memory, and is read by the reliability calculator 3 at the time of calculating the next reliability. In the figure, reference numeral 5 denotes a reliability memory. 4 is
A smoother that calculates a smoothed value and a smoothed error covariance based on the above-described observation value, reliability of each motion model, a constant acceleration vector, an acceleration smoothed value, a gain matrix, and a smoothed value and a smoothed error covariance matrix before one sampling. The smoothed value and smoothed error covariance matrix before one sampling are accumulated in the smoother memory, and are read by the smoother 4 when calculating the next smoothed value and smoothed error covariance. In the figure, reference numeral 6 denotes a smoother memory.

[0005] Reference numeral 7 denotes a predictor for calculating a prediction value and a prediction error covariance matrix based on the smoothed value calculated by the smoother 4, each motion model reliability, and a constant acceleration vector. Is a first delay circuit for delaying the predicted position calculated by the above by a unit time.

Reference numeral 9 denotes a predictor for each motion model which calculates a predicted value for each motion model based on the above-mentioned smoothed value. Reference numeral 10 denotes an observation error covariance matrix obtained from a preset observation model. And a gain matrix calculator that calculates a gain matrix used for calculating a smoothed value, based on a prediction error covariance matrix obtained from the predictor 9 for each motion model calculated one sampling before.

Reference numeral 11 denotes a model setting device for setting a constant acceleration vector and a model transition matrix. The set constant acceleration vector is input to the smoother 4, the predictor 7, the predictor 9 for each model, and the input vector smoother 12, and the model transition matrix is input to the reliability calculator 3.

Reference numeral 12 denotes an input vector smoother for calculating an input vector smoothed value from the calculated reliability of each model and a set constant vector. Reference numeral 13 denotes an input vector smoother based on the calculated reliability of each model and a model transition matrix. , An input vector, and an input vector predictor for calculating a predicted value. Reference numeral 14 denotes a second delay circuit that delays the prior reliability of each model by a unit time. Reference numeral 15 denotes a display device that discretely displays the smoothed values on a screen.

Next, the operation of the above configuration will be described with reference to the flowchart of FIG. In step 13-1, a smoothed value of the target position / velocity and an initial value of the smoothed error covariance matrix are set based on the target Kalman filter theory based on the target position observation information obtained from the target observation device 1. .

Next, in step 13-2, in each motion model predictor 9, N number of constant acceleration vectors (including zero acceleration vector) as shown in FIG. An exercise model is set.

In step 13-3, in the predictor 9 for each motion model, the smoothed value input from the smoother 4 and the constant acceleration vectors constituting the N motion models are used for each motion model. A prediction value is calculated, and a prediction error covariance matrix for each motion model is calculated using the smoothing error covariance matrix input from the smoother 4 and a preset driving noise covariance matrix.

Next, at step 13-4, the predictor 7
, The smoothed value and smoothed error covariance matrix output from the smoother 4, the reliability of each motion model output from the reliability calculator 3, and the input vector predicted value output from the input vector predictor 13 are read. It is. The predictor 7 calculates a predicted value and a prediction error covariance matrix based on the input data.

Next, at step 13-5, the target observation device 1 reads target position observation information. The above observation value is a result of detection of a signal from a target and selection of a tracking target by gate determination, and is constituted by position information.

Next, in step 13-6, the observation value obtained by the gate decision unit 2 is output to the reliability calculation unit 3. The reliability calculator 3 further reads the reliability of each motion model calculated one sample before the current time stored in the reliability memory 5, and the predictor 9 for each motion model calculates the motion calculated one sample before. The prediction value and the prediction error covariance matrix for each model are read via the fourth delay circuit 16. The reliability calculator 3 receives the input data, the observation error covariance matrix, and the model setting unit 11.
The reliability of the exercise model is calculated based on the transition probabilities of the exercise model input from.

Next, at step 13-7, the gain matrix calculator 10 calculates, via the third delay circuit 17, the prediction error covariance matrix calculated by the predictor 9 for each motion model one sampling before the current time. Read. The gain matrix calculator 10 calculates a gain matrix based on the input data and a preset observation error covariance matrix.

Next, in step 13-8, the smoothing unit 4
, The smoothed value and the smoothed error covariance matrix calculated one sampling before the current time are read from the smoothing memory 6. The smoother 4 further includes a reliability of each motion model output by the reliability calculator 3, a gain matrix output by the gain matrix calculator 10, an input vector smoothed value output by the input vector predictor 13, and a model setting unit 11. Reads the constant acceleration vector and observation value output by. The smoother 4 is
A smoothed value and a smoothed error covariance matrix are calculated based on the input data.

Next, at step 13-9, it is determined whether or not the above-described series of processing has been completed and the end of the tracking processing has been requested. As a result, if it is determined that the end of the process has not been requested, the process is repeated.

As described above, the conventional target tracking device models a target motion using a plurality of motion models, and the expected existence probability of the target centered on the predicted position of each motion model is a three-variable model. According to the normal distribution, the equiprobability line becomes an error ellipsoid as shown in FIG. 15A when there are five motion models. If the target exists inside the error ellipsoid, the motion of the target can be represented using a plurality of motion models.

[0019]

In the conventional target tracking apparatus, the flatness of the above-mentioned error ellipsoid is determined by the sum of the observation error covariance matrix and the prediction error covariance matrix. In the polar coordinate system, the observation noise is assumed to follow white Gaussian noise, but since the observation noise needs to be converted to a rectangular coordinate system, the oblateness of the error ellipsoid increases as the angle error is greater than the distance error . Therefore, when the angle error is larger than the distance error and the sampling interval is larger, FIG.
As shown in (b), the error ellipsoid is flattened, and the predicted positions between the motion models are separated, and the tracking accuracy with respect to the turning target is likely to be deteriorated particularly with the expansion of the area where the motion of the target cannot be described. was there. In addition, since the target is assumed to continue turning for one sampling period, when the sampling interval is large and the turning duration of the target is short, the change in the position and speed of the target is estimated to be larger than the actual one. There is a problem that accuracy is easily deteriorated.

The present invention has been made in order to solve the above-mentioned problems, and the distance, the center distance, and the number of models of the error ellipsoid are controlled so that the error ellipsoids always overlap each other. To provide a target tracking device capable of obtaining good tracking accuracy even in a severe environment in which a distance target is tracked at a large sampling interval, and to prevent a position / velocity occurring when the target turning duration is shorter than the sampling interval. An object is to provide a target tracking device that suppresses necessary compensation.

[0021]

A target tracking device according to the present invention uses a Kalman filter for N motion models composed of N constant acceleration vectors.
Suppresses the oscillation phenomenon in which the motion model is switched at each sampling due to the extremely high reliability of the specific motion model and the accompanying deterioration of tracking accuracy, and the uniformity of the reliability of each motion model and the deterioration of observation noise reduction performance In the target tracking device provided with a control device that performs the above, based on the prediction value and the prediction error covariance matrix for each of the N motion models calculated by the Kalman filter, the target tracking with respect to the prediction positions of the N motion models is performed. Error ellipsoid overlap degree determination means for determining the overlap degree of the error ellipsoid which is an equal probability line indicating the expected existence probability, based on the overlap degree of the error ellipsoid determined by the error ellipsoid overlap degree determination means, An error ellipsoid enlarging / reducing means for enlarging and reducing the error ellipsoid so that no blank area occurs between the error ellipsoids.

The error ellipsoid enlarging / reducing means includes:
The error ellipsoid is enlarged and reduced by controlling the number of target motion models and the magnitude of the constant acceleration vector.

Further, by using a Kalman filter for N motion models composed of N constant acceleration vectors, the reliability of a specific motion model becomes extremely high, and the oscillation phenomenon in which the motion model is switched for each sampling and the oscillation phenomenon. In the target tracking device equipped with a control device that suppresses the accompanying degradation of tracking accuracy and the uniformity of reliability of each motion model and the resulting degradation of observation noise reduction performance, the above Kalman filter keeps turning the target during one sampling. The target turn duration time is estimated from the target model information observed by the target observation device based on the above assumption, and the target maximum acceleration performance, and the change in speed or acceleration due to the turn is excessive. Turning duration determination means for determining the possibility of calculation, and external input in a calculation formula for calculating the reliability. It is obtained by a damping gain calculating means for calculating an attenuation gain to attenuate the effects of terms.

Further, the error ellipsoid overlap degree judging means includes:
The ratio of the major axis / minor axis / medium diameter of the error ellipsoid of each motion model to the Euclidean distance, which is the distance between the predicted position of each motion model and the constant velocity motion model, is compared with a preset threshold value. By doing so, the degree of overlap of the error ellipsoid is determined.

Further, the error ellipsoid overlap degree determining means includes:
By comparing the residual quadratic form in which the Euclidean distance is normalized with the residual error covariance matrix in the calculation formula for calculating the reliability, and a preset threshold value,
This is for determining the degree of overlap of the error ellipsoid.

The turning duration determining means compares the ratio between the determined turning duration of the target and the sampling interval with a predetermined threshold value, whereby the change in speed or acceleration due to turning is excessively calculated. Is to be determined.

The major axis of the error ellipsoid of each motion model
Based on the ratio of the short diameter / medium diameter and the Euclidean distance that is the distance between the predicted position of each of the motion models and the constant velocity motion model,
The apparatus further comprises a multiplication coefficient calculation means for calculating a multiplication coefficient of the error covariance matrix in the calculation formula for calculating the reliability.

The multiplication coefficient calculating means uses a residual quadratic form obtained by normalizing a distance between an observation value observed by the target observation device and a predicted position of each motion model by a residual error covariance matrix. , The error ellipsoid is enlarged and reduced.

Further, using the ratio of the major axis, minor axis, and medium axis of the calculated error ellipsoid of the motion model to the Euclidean distance that is the distance between the predicted position of each motion model and the constant velocity motion model, It is provided with model control means for recalculating the constant acceleration vector of each motion model and adjusting the predicted position between the motion models in which the error ellipsoid does not overlap.

Further, by using the ratio of the major axis, minor axis, and medium axis of the calculated error ellipsoid of the motion model to the Euclidean distance, which is the distance between each motion model and the predicted position of the constant velocity motion model, It is provided with a model control means for specifying a region where the error ellipsoids of each motion model do not overlap and adding a new motion model to the region.

Further, the damping gain calculating means calculates the damping gain of the constant acceleration vector based on the sampling interval and the estimated turning duration of the target.

[0032]

DESCRIPTION OF THE PREFERRED EMBODIMENTS Embodiment 1 Hereinafter, a Kalman filter using a multiple motion model having N motion models used as a tracking filter in the target tracking device according to the present invention will be described.

A target motion model is defined as in equation (1).

(Equation 1) Here, X _k is a state vector representing the true value of the target motion data at the sampling time t _k , for example, xy
The target position vector in the rectangular coordinate system of z is expressed by equation (2),
When the target velocity vector in the orthogonal coordinate system is represented by Expression (3), X _k is _represented by Expression (4).

(Equation 2)

Φ _k-1 is a transition matrix representing the transition of the state vector X _k from the sampling time t _k-1 to t _k . Equation (5) is based on the assumption that the target performs a uniform linear motion. . Here, I is a unit matrix as shown in Expression (6).

(Equation 3)

[0035] w _k in the drive noise vector at the sampling time t _k, is Γ _k is a transformation matrix of the drive noise vector at the sampling time t _k. The term of the truncation error due to the assumption that the target motion model is a uniform linear motion is Γ _k-1
Assuming that w _k−1 , w _k is equivalent to an acceleration vector, and Γ
_k-1 is the equation (7).

(Equation 4)

W _k is a three-dimensional normally distributed white noise having an average of 0, and is represented by equations (8) and (9). Here, 0 represents a zero vector, and Q _k is a driving noise covariance matrix at the sampling time t _k .

(Equation 5)

U _k is a constant acceleration vector constituting the N motion models at the sampling time t _k , and is represented by equation (10). Gamma 'is at the sampling time t _k, the transform matrix of constant acceleration vector is an expression (11).

(Equation 6)

FIG. 14 is a diagram for explaining the constant acceleration vector. In the figure, O is the origin of coordinates O-xyz with the origin of the target observation device, and X is the coordinates O-x with the east direction being the positive x-axis. The x-axis of xyz, Y is the coordinate O where the north direction is the positive y-axis, and the Y-axis of x-xyz, and Z is the coordinate O where the upward is the positive z-axis
−xyz z-axis, α ₁ is a constant acceleration vector in the y-axis positive direction, α ₂ is a constant acceleration vector in the y-axis negative direction, α ₃ is a constant vector in the X-axis positive direction, α ₄ is a constant in the X-axis positive direction Acceleration vector, α ₅ is a constant acceleration vector in the positive direction of the Z-axis, α ₆
Is a constant acceleration vector in the z-axis negative direction. In addition,
In the case of the motion model considering the constant acceleration vector α ₇ = 0, the number of models N is seven. In one axis of constant acceleration,
North direction of north reference Cartesian coordinate system, estimated velocity vector, LOS
(Line OfSight). Here, the LOS is a vector connecting the radar and the target.

Next, a hypothesis that Expression (12) is true at the sampling time t _k is written as Expression (13).

(Equation 7)

Assume Markov nature in the movement of the motion model. That is, the motion model Ψ _{k, α} is the sampling time t
_It is determined by the motion model of _k-1 and does not depend on the motion model up to the sampling time t _k-2 .

[0041] between the motion model at time t _k the transition probability represented by the equation (14). For example, the minimum sampling interval is t _s, p a motion model between transition probability when the sampling interval _{t k -t k-1 = t} s in the same meaning as in the formula (14)
_ab . Further, an element of a row and b column in an N × N matrix is defined as p _ab
Is expressed as a transition probability matrix between motion models Π. Sampling interval _{t k -t k-1 = m} · t s, (m = 1,2,
..), The transition probability matrix 運動_k between the motion models at the sampling time t _k can be _represented by Expression (15). At this time, the element of a row and b column of Π _{k is} the transition probability between motion models p _{k, ab}
It is.

(Equation 8)

[0042] Based on the observation information Z ^k to the sampling time t _k, the conditional probability density function the reliability of the motion model at the sampling time t _k, by defining the formula (16), is calculated by the equation (17) You.

(Equation 9) Here, ν _{k, a} is a normal distribution approximation P [Z _k
| Ψ _{k, a,} which was approximated by a multivariate normal distribution Z ^k-1],
It is represented by equation (18).

(Equation 10)

Observation information Z up to sampling time t _k-1
^If the prior reliability of the motion model at the sampling time t _k based on ^k−1 is defined by the conditional probability density function as Expression (19), Expression (20) can be written.

[Equation 11]

[0044] Based on the observation information Z ^k to the sampling time t _k, the estimated value of the constant acceleration vector is defined as equation (21), written as Equation (22). This is called an estimated acceleration vector.

(Equation 12)

Observation information Z up to sampling time t _k-1
The estimated value of the constant acceleration vector based on ^k-1 is expressed by equation (2)
If defined as 3), it can be written as equation (24). This is called a predicted acceleration vector.

(Equation 13)

It is assumed that one observation value from the target to be tracked is obtained at the sampling time t _k , and the observation system model is represented by Expression (25). Here, Z _k is the position measurement vector according constituted orthogonal coordinate than the observed value of the position information at sampling time t _k, H is the observation matrix at sampling time t _k, a formula (26). v _k is the observation noise vector corresponding to the observation vector Z _k at the sampling time t _k, a three-dimensional normal distribution white noise of zero mean, a formula (27), equation (28). Note that R _k is an observation error covariance matrix at the sampling time t _k and is a value that does not depend on the motion model. Further, it is assumed that the driving noise vector and the observation noise vector are independent of each other.

[Equation 14]

Further, the accumulation of the observation value vectors from the sampling time t ₁ to t _k is represented by the equation (29).

(Equation 15)

According to the Kalman filter theory, the estimated value X _k (+) of the state vector X _k when the observed value is obtained at the sampling time t _k according to the above model is expressed by the following equation (30).
３８ (38).

(Equation 16)

Here, X _{k, a} (−), X _{k, a} (+), X _{k
(-), X k (+} ), P k, a (-), P k, a (+), P
_k (−) and P _k (+) are calculated by equations (39) to (46), respectively.
Is defined as

[Equation 17]

[0050] X _{k, a} (-) is X _k based on the hypothesis Ψ _{k, a} of observation information Z ^k-1 and the motion model of up to sampling time t _k-1
In conditional mean value of corresponds to the prediction vector estimating the true value of hypothetical [psi _{k, a} observation information Z ^k-1 and motion model to the sampling time t _k-1 at the sampling time t _k.

[0051] X _{k, a} (+) in the conditional mean value of X _k based on the hypothesis Ψ _{k, a} of the motion model observation information Z ^k up to the sampling time t _k, observation information Z of up to sampling time t _k ^This corresponds to a smoothed vector whose true value has been estimated at the sampling time t _k based on _{k and} the hypothesis 運動_{k, a} of the motion model.

[0052] X _k (-) is a conditional mean value of X _k based on the observation information Z ^k-1 to the sampling time t _k-1, based on the observation information Z ^k-1 to the sampling time t _k-1 Te corresponds to the prediction vector estimating the true value at the sampling time t _k.

X_k(+) Indicates sampling time t_kUp to
Measurement information Z^kX based on_kThe conditional average of the sample
Time t_kObservation information Z up to^kWhen sampling based on
Time t _kCorresponds to the smoothed vector whose true value has been estimated.

P _{k, a} (−), P _{k, a} (+), P _k (−),
P _k (+) is X _{k, a} (−), X _{k, a} (+), X _k
(−), A prediction error covariance matrix for each motion model representing an error covariance matrix of X _k (+), a smooth error covariance matrix for each motion model, a prediction error covariance matrix for N motion models,
And a smooth error covariance matrix with N motion models.

The formula (34) is a gain matrix at sampling time t _k.

It is assumed that the initial values X ₀ (+) and P ₀ (+) are separately determined in the same manner as in the case where the Kalman filter is normally applied. From equation (31), P _{k, a}
Since (−) is _a value that does not depend on the hypothesis Ψ _{k, a} , K _k from Equation (34) and P _{k, a} (+) from Equation (36) similarly have the hypothesis Ψ _{k, a}
The value does not depend on

Next, FIG. 1 is a block diagram of the target tracking device according to the first embodiment. 1, the same reference numerals as those in FIG. 12 indicate the same parts, and the description thereof will be omitted. As a new code, 18 is an error ellipsoid overlap determiner, 19 is an error covariance matrix controller, and 20 is a multiplier.

The error ellipsoid overlap determiner 18 obtains the prediction value for each motion model and the prediction error covariance matrix for each motion model from the predictor 9 for each model, and uses Output to the variance matrix controller 19.

Further, the error covariance matrix controller 19 outputs the prediction value for each motion model, the prediction error covariance matrix for each motion model, and the prediction error covariance matrix for each motion model from the prediction unit 9 for each model via the error ellipsoid overlap determination unit 18. The overlap determination information is acquired from the error ellipsoid overlap determiner 18, and when the error ellipsoid overlap determiner 18 determines that the degree of overlap of the error ellipsoid is not appropriate, the error ellipsoid enlargement / reduction is performed based on the information. A multiplication coefficient for determining the reduction is calculated.

The multiplier 20 multiplies the multiplication coefficient obtained from the error covariance matrix controller 19 by the residual error covariance matrix used for the reliability calculation, and outputs the result to the reliability calculator 3.

FIG. 2 is a flowchart showing the operation of the error ellipsoid duplication determiner 18 according to the first embodiment, which will be described below with reference to the drawings.

First, in step 2-1, the error ellipsoid duplication determiner 18 calculates the equation (3) from the predictor 9 for each model.
0) and the prediction error covariance matrix for each motion model shown in equation (31) are obtained, and the predicted positions of the constant velocity motion model and another motion model, that is, the centers of the error ellipsoids are obtained. Calculate the Euclidean distance of. The Euclidean distance between the predicted positions is represented by Expression (47).

(Equation 18) Here, HX _{k, 1} (−) is the predicted position of the constant velocity motion model.

Next, in step 2-2, the major axis, middle axis, and minor axis of the error ellipsoid are obtained from equation (48).

[Equation 19] Here, λ _{k, a, i} is an eigenvalue of the residual error covariance matrix S _{k, a} and determines the size of the error ellipsoid.

Next, in step 2-3, the ratio γ _{k, a} of the diameter length / center distance is calculated from the equation (49) for the vector connecting the centers which are in parallel with the respective diameters.

(Equation 20)

Next, in step 2-4, each
For the motion model, the calculated ratio γ _{k, a}Two thresholds
Value, and γ_{k, a}Is smaller,
A constant coefficient ε = 1 is assigned, and γ_{k, a}But
If it is larger, a decision coefficient ε = 2 is assigned,
Is not applicable, ε = 0 is assigned. However
However, the threshold value 1 is smaller than the threshold value 2 and ε
= 1 means that the error ellipsoids are not sufficiently overlapped, and ε =
2 indicates that the error ellipsoids are too overlapped, and ε = 0 indicates an error.
This indicates that the difference ellipsoids are appropriately overlapped.

FIG. 3 is a flow chart showing the operation of the error covariance matrix controller 19 and the multiplier 20 using the diameter of the error ellipsoid and the center distance ratio of each motion model in the first embodiment. This will be described below with reference to the drawings.

First, in step 3-1, the error covariance matrix controller 19 acquires the judgment information from the error ellipsoid overlap judgment unit 18, and the error ellipsoids are appropriately overlapped (ε
= 0). If it is determined that ε = 0, the process proceeds to step 3-6, and it is determined whether or not to end for all the motion models. If it is determined that the error ellipses do not sufficiently overlap (ε = 1) or that they overlap too much (ε = 2), the process proceeds to step 3-2.

Next, from step 3-3 to step 3-
In step 5, the major axis, middle axis, and minor axis of the error ellipsoid are calculated from equation (50), and the distance between the predicted positions of the constant velocity motion model and another motion model is calculated from equation (47).

As a result of the above processing, in step 3-6,
It is determined whether or not the processing has been completed for all exercise models. If the processing is not completed, steps 3-1 to 3-5 are repeated, and if the processing is completed, the process proceeds to step 3-7.

Next, in step 3-7, the ratio γ _{k, a} of the distance of the vector connecting the prediction center parallel to the major axis, middle axis, and minor axis of the error ellipsoid is calculated from equation (51), and each motion is calculated. Calculate the minimum value of γ _{k, a} of the model.

Next, in step 3-8, since the case where the error ellipsoids do not overlap has a remarkable effect on the tracking accuracy, the minimum value of the ratio and the threshold value 1 are compared first, and the minimum value is determined by the threshold value. If it is less than 1, a multiplication coefficient for enlarging the error ellipsoid is calculated. The multiplication coefficient is calculated such that γ _{k, a} has the same value as threshold value 1. On the other hand, when the minimum value does not fall below the threshold value 1, the minimum value is compared with the threshold value 2. When the minimum value exceeds the threshold value 2, the multiplication coefficient for reducing the error ellipsoid is calculated. Is calculated. The multiplication coefficient is calculated so that γ _{k, a} has the same value as threshold value 2. Here, the threshold value 2 is larger than the threshold value 1.

In step 3-9, the multiplication coefficient calculated by the error covariance matrix controller 19 in the above step is input to the multiplier 19, and the multiplier 19 calculates the equation (18).
Is multiplied by the residual error covariance matrix in the reliability calculation of.

As described above, according to the target tracking device according to the first embodiment, the possibility of reliability oscillation is known in advance by judging the degree of overlap of the error ellipsoid, and the error ellipse of each motion model is determined. By enlarging / reducing the body, the error ellipsoids can always be overlapped, and the reliability can be mixed. Therefore, the desired tracking accuracy with respect to the turning target can be maintained even under conditions where the reliability is likely to oscillate, such as when the sampling interval is long and a long distance target is tracked.

Embodiment 2 FIGS. 4 and 5 are flowcharts showing the same configuration as the above-described target tracking device according to the first embodiment, and showing an operation different from that of the first embodiment.

FIG. 4 shows the operation of the error ellipsoid overlap determiner 18 of the target tracking device according to the second embodiment.
In the figure, in step 4-1, the error ellipsoid overlap determiner 18 obtains a prediction value for each motion model shown in equation (30) and a prediction error for each motion model shown in equation (31) from each model predictor 9. The covariance matrix is obtained, the observation values are further obtained, and the residual quadratic form shown in equation (50) for each motion model is calculated.

(Equation 21) Here, S _{k, a} is a residual error covariance matrix. Equation (5
A region where the target is expected to exist with a certain probability satisfying 0) is an error ellipsoid.

Since the residual follows a trivariate normal distribution,
The residual quadratic form follows a chi-square distribution. Therefore, the threshold d can be obtained from the χ-square distribution by determining the existence probability of the target.

Next, at step 4-2, the minimum value of the residual quadratic form of each motion model is compared with the threshold value 1. If the minimum value is larger than the threshold value 1, the overlap of the error ellipsoids is determined. When it is determined that the error ellipsoids are not enough, the determination coefficient ε = 1 is assigned, and the variance of the residual quadratic form of each motion model is smaller than the threshold value 2, it is determined that the error ellipsoids overlap each other too much. It is determined and a determination coefficient ε = 2 is assigned. If neither of the above cases applies, it is determined that the error ellipsoids are appropriately overlapped, and the determination coefficient ε = 0 is assigned. However, the threshold value 1 is a value smaller than the threshold value 2.

Then, in step 4-3, step 4
The determination information including a table of the determination coefficients assigned according to -2 is created.

FIG. 5 shows an error co-distribution matrix controller 19 based on the determination information created by the error ellipsoid overlap determiner 18 of the target tracking apparatus according to the second embodiment.
And the operation of the multiplier 20. In the figure, in step 5-1, the error covariance matrix controller 19
Obtains determination information from the error ellipsoid overlap determiner 18,
It is determined whether or not the error ellipsoids are appropriately overlapped (ε = 0). When it is determined that ε = 0, the processing here ends, and when it is determined that the error ellipses do not sufficiently overlap (ε = 1) or that the error ellipses overlap too much (ε = 2), Proceed to step 5-2.

Next, in step 5-2, the residual quadratic form of the predicted position of each motion model and the observed value is calculated from equation (48).

Next, in step 5-3, when the overlap of the error ellipses is not sufficient (ε = 1), a multiplication coefficient is calculated so that the minimum value of the residual quadratic form becomes equal to the threshold value 1. On the other hand, if it is determined that the error ellipses are too overlapped (ε = 2), a multiplication coefficient is calculated such that the maximum value of the residual quadratic form becomes equal to the threshold value 2.

Then, in step 5-4, the multiplier 20
Multiplies the multiplication coefficient obtained from the error covariance matrix controller 19 by the residual error covariance matrix in the reliability calculation of Expression (18).

As described above, according to the target tracking apparatus according to the second embodiment, for the N motion models, the Euclidean distance between the observed value and the predicted positions of all the motion models is used as the residual error of the filter. By comparing the residual quadratic form normalized by the variance matrix with the threshold value, it is determined whether the error ellipsoids are appropriately overlapped, and using the residual quadratic form of each motion model, By calculating a multiplication coefficient for enlarging and reducing the error ellipsoid and enlarging and reducing the error ellipsoid, the reliability can be appropriately mixed. As a result, a desired tracking accuracy with respect to the turning target can be maintained even under conditions where the reliability is likely to oscillate, such as when the sampling interval is long and a long distance target is tracked.

Embodiment 3 FIG. 6 is a block diagram of a target tracking device according to the third embodiment having a configuration different from those of the first and second embodiments. In FIG. 6, 21 is a model controller as a new code.

The error ellipsoid overlap determiner 18 obtains the prediction value for each motion model and the prediction error covariance matrix for each motion model from the predictor 9 for each model, and performs model control on the error degree ellipsoid overlap degree determination information. Output to the container. Also, the model controller 21
Is based on the prediction value for each motion model obtained from the prediction unit 9 for each model, the prediction error covariance matrix for each motion model, and the overlap determination information obtained from the error ellipsoid overlap determination unit 18. When the overlap determiner 18 determines that the degree of overlap of the error ellipsoid is not appropriate, the number of models is increased and the constant acceleration vector is recalculated based on the information,
These pieces of information are output to the model setting device 11.

Next, the operation of the target tracking device according to the third embodiment will be described. Note that the error ellipsoid overlap determiner 18 in this configuration is the same as that in the first or second embodiment described above.
Since the same operation (FIG. 2 or FIG. 4) is performed, the description thereof is omitted. FIG. 7 shows a model controller 2 according to the third embodiment.
2 is a flowchart showing the operation of the first embodiment, which will be described below using the drawings.

In the figure, in step 7-1, the model controller 21 acquires the judgment information from the error ellipsoid overlap judgment unit 18, and the error ellipsoids are appropriately overlapped (ε = 0).
It is determined whether or not. If it is determined that ε = 0, the process proceeds to step 7-7. The overlap of the error ellipses is not enough (ε
= 1) or too much overlap (ε = 2), the process proceeds to step 7-2.

Next, in step 7-2, the major axis, the medium axis, and the minor axis of the error ellipsoid are expressed by equation (50), and the distance between the constant velocity motion model and the predicted position of another motion model is expressed by equation (47). It is calculated from:

Next, from step 7-3 to step 7-
In step 5, the ratio γ _{k, a} of the distance of the vector connecting the prediction centers parallel to the major axis, middle axis, and minor axis of the error ellipsoid is calculated from equation (51).

Next, in step 7-6, a constant acceleration vector of each motion model is calculated so that the distance ratio γ _{k, a} of each motion model becomes equal to the threshold value.
Output to the model setting device 11.

As a result of the above processing, in step 7-7,
It is determined whether or not the processing for all the exercise models has been completed. If the processing has not been completed, steps 7-1 to 7-6 are repeated, and the constant acceleration vector of each motion model is recalculated.

As described above, according to the third embodiment,
Using the ratio of the calculated major axis / minor axis / medium axis of the error ellipsoid of the motion model and the Euclidean distance that is the distance between the predicted position of each motion model and the constant velocity motion model, the constant of each motion model is calculated. By providing the model control means for recalculating the acceleration vector and adjusting the predicted position between the motion models where the error ellipsoid does not overlap, the center of the error ellipsoid is moved to appropriately overlap the error ellipsoid, and the reliability is improved. Can be mixed. As a result, a desired tracking accuracy with respect to the turning target can be maintained even under conditions where the reliability is likely to oscillate, such as when the sampling interval is long and a long distance target is tracked.

Embodiment 4 FIG. 8 is a flowchart showing the operation of the model controller 21 having the same configuration as the above-described target tracking device according to the third embodiment and different from the third embodiment. This will be described below with reference to the drawings.

The operations in Steps 8-1 to 8-5 are the same as those in Steps 7-1 to 7-5 of the flowchart in Embodiment 3 described above, and thus the description will be omitted.

In steps 8-6 and 8-7, the distance ratio γ _{k, a of} each motion model obtained in step 8-5 is calculated as
When the error ellipsoids are far apart from each other, a new motion model is added, and its initial value and a constant acceleration vector are calculated.

As a result of the above processing, in step 8-8,
It is determined whether the processing for all the exercise models has been completed. If the processing has not been completed, steps 8-1 to 8-7 are repeated.

Then, in step 8-9, when the addition of the new motion model is completed, the model transition probability shown in the equation (15) is calculated, and this is input to the model setting unit 11.

As described above, according to the fourth embodiment,
The model control means uses the ratio of the calculated major axis / minor axis / medium axis of the error ellipsoid of the motion model to the Euclidean distance that is the distance between each motion model and the predicted position of the constant velocity motion model. By specifying an area where the error ellipsoids of each motion model do not overlap and adding a new motion model to the area, the reliability of each motion model can be mixed appropriately.

Embodiment 5 FIG. FIG. 9 is a block diagram of a target tracking device according to the fifth embodiment having a configuration different from those of the first to fourth embodiments. In FIG. 9, as a new code, 22 is a turning duration determination unit, and 23 is a damping gain calculator.

The turning duration determination unit 22 obtains the smoothing speed, the target model information, and the target maximum acceleration performance from the smoothing unit 4, and estimates the turning duration. By comparing the turning duration time with the sampling interval, it is determined that the position or speed change may be excessively calculated due to model mismatch, and the determination information is output to the attenuation gain calculator 23.

The damping gain calculator 23 calculates a damping gain based on the judgment information obtained from the turning duration determining unit 22 and outputs the obtained information to the model setting unit 11.

The model setting unit 11 calculates the obtained attenuation gain by the equation (1) in the above-described Kalman filter calculation equation.
Multiply the constant acceleration conversion matrix Γ ′ shown in 1).

FIG. 10 is a flowchart showing the operation of the turning time determiner 22 in the fifth embodiment.
This will be described below with reference to the drawings.

First, in step 6-1, the turning duration determination unit 22 determines the target model based on the smoothing speed vector obtained from the smoothing unit 4 and the sampling interval, and estimates the target turning duration.

Next, in step 6-2, equations (52) to (54) show relational expressions between the turning cycle, the speed, and the turning load. Here, the turning cycle is the time required for the target to make a 360 ° turn on the same plane at a constant flight speed and turning load. If the target speed and the turning load are obtained, the turning cycle, The turning duration can be estimated.

(Equation 22) Here, T is the turning cycle, r is the turning radius, v is the flight speed,
g is the gravitational acceleration, θ is the inclination angle (bank angle) of the plane of symmetry of the body with respect to the vertical plane, and n is the load multiple.

Next, at step 6-3, if the estimated turning duration of the target is significantly shorter than the sampling interval, it is determined that the attenuation gain needs to be inserted, and the determination information is output to the attenuation gain calculator 23. .

FIG. 11 is a flowchart showing the operation of the attenuation gain calculator 23 and the model setting unit 11 according to the fifth embodiment, which will be described below with reference to the drawings.

In FIG. 11, in step 11-1, the damping gain calculator 23 calculates a damping gain based on the possibility of execution of damping gain insertion obtained from the turning duration determiner 22 and the sampling interval and the estimated turning duration. I do.
Here, the attenuation gain is a diagonal matrix as shown in Expression (55), and the coefficients ζ and η are the attenuation gains of the external input terms with respect to position and velocity, respectively.

(Equation 23)

Next, at step 11-2, the model setting unit 11 calculates the equations (30), (32), (33), (3)
7) Multiply Γ _k 'shown in (38) by the attenuation gain.

As described above, according to the fifth embodiment,
The damping gain calculator calculates the damping gain of the constant acceleration vector on the basis of the sampling interval and the estimated target turning duration, so that the influence of the external input term can be damped. This is caused when the turning duration is significantly shorter than the sampling interval. It is possible to suppress a remarkable increase in the position or the speed due to the model mismatch, and the deterioration of the tracking accuracy accompanying the increase.

[0111]

As described above, according to the target tracking apparatus of the present invention, the possibility of reliability oscillation is known in advance by judging the degree of overlap of the error ellipsoid, and the error ellipsoid of each motion model is determined. The error ellipsoid can be appropriately overlapped by enlarging / reducing, and the reliability can be mixed. This allows
Even under conditions where the sampling interval is long and tracking of a long distance target is likely to occur, it is possible to maintain a desired tracking accuracy for the turning target while reducing errors due to observation noise.

Further, by judging the degree of overlap of the error ellipsoids, it is known in advance that the tracking accuracy may be degraded due to the oscillation of the reliability. By controlling the magnitude of the vector, the error ellipsoids of each motion model can be appropriately overlapped and the reliability can be mixed appropriately.

Further, by estimating the turning duration from the target model, speed, and the target maximum acceleration performance, it is possible to know in advance that the model does not match due to a marked difference between the target turning duration and the sampling interval. it can. Also, since the Kalman filter using the multiple motion model assumes that the target continues turning during one sampling, if the estimated turning continuation time is smaller than the sampling interval, a damping gain for attenuating the influence of the external input term is calculated. Thus, it is possible to suppress a remarkable increase in position or speed due to model mismatch and deterioration of tracking accuracy due to this.

The error ellipsoid overlap degree judging means calculates the Euclidean distance, which is the distance between the major axis, minor axis, and medium axis of the error ellipsoid of each motion model and the predicted position of each motion model and the constant velocity motion model. By comparing this ratio with a preset threshold value, it can be determined whether or not the error ellipsoids overlap appropriately, and the possibility of reliability oscillation can be known in advance.

The error ellipsoid overlap degree judging means includes a residual quadratic form obtained by normalizing the Euclidean distance with the residual error covariance matrix in the calculation formula for calculating the reliability, and a predetermined threshold. By comparing the values with the values, it is possible to determine whether the error ellipsoids are appropriately overlapped, and to know in advance the possibility of reliability oscillation.

The turning duration determining means compares the ratio between the determined turning duration of the target and the sampling interval with a predetermined threshold value, so that the change in speed or acceleration due to turning is excessively calculated. In order to determine the possibility, it is possible to know in advance the possibility of a remarkable increase in the estimated speed vector due to model mismatch and the deterioration of the tracking accuracy.

The major axis of the error ellipsoid of each motion model
Based on the ratio of the short diameter / medium diameter and the Euclidean distance that is the distance between the predicted position of each of the motion models and the constant velocity motion model,
By further comprising a multiplication coefficient calculating means for calculating a multiplication coefficient of the error covariance matrix in the calculation formula for calculating the reliability, the reliability is appropriately increased by enlarging / reducing the error ellipsoid of each motion model. Can be mixed.

Further, the multiplication coefficient calculating means uses a residual quadratic form obtained by normalizing a distance between an observation value observed by the target observation device and a predicted position of each motion model by a residual error covariance matrix. Since the error ellipsoid is enlarged and reduced, the reliability can be appropriately mixed. As a result, even when the sampling interval is long and tracking of a long-distance target is likely to occur, even under conditions where oscillation of reliability is likely to occur,
A desired tracking accuracy for the turning target can be maintained.

Further, using the ratio of the major axis, minor axis, and medium axis of the calculated error ellipsoid of the motion model to the Euclidean distance, which is the distance between the predicted position of each motion model and the constant velocity motion model, By providing a model control means for recalculating the constant acceleration vector of each motion model and adjusting a predicted position between the motion models in which the error ellipsoid does not overlap, the center of the error ellipsoid is moved and the error ellipsoid is appropriately adjusted. And reliability can be mixed. As a result, a desired tracking accuracy with respect to the turning target can be maintained even under conditions where the reliability is likely to oscillate, such as when the sampling interval is long and a long distance target is tracked.

Further, by using the ratio of the major axis, minor axis, and medium axis of the calculated error ellipsoid of the motion model to the Euclidean distance, which is the distance between the predicted position of each motion model and the constant velocity motion model, By specifying a region where the error ellipsoids of each motion model do not overlap and by providing a model control unit for newly adding a motion model to the region, the reliability of each motion model can be appropriately mixed.

Further, the damping gain calculating means calculates the damping gain of the constant acceleration vector based on the sampling interval and the estimated target turning duration, thereby attenuating the influence of the external input term. it can. This is caused when the turning duration is significantly shorter than the sampling interval. It is possible to suppress a remarkable increase in the position or the speed due to the model mismatch, and the deterioration of the tracking accuracy accompanying the increase.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a configuration of a target tracking device according to Embodiment 1 of the present invention.

FIG. 2 is a flowchart showing an operation of an error ellipsoid overlap determination unit 18 of the target tracking device according to the first embodiment of the present invention.

FIG. 3 is a flowchart illustrating operations of an error covariance matrix controller and a multiplier using a diameter of an error ellipsoid and a center distance ratio of each motion model of the target tracking device according to the first embodiment of the present invention;

FIG. 4 illustrates an operation of an error ellipsoid overlap determiner of the target tracking device according to the second embodiment of the present invention.

FIG. 5 is a flowchart showing operations of an error co-distribution matrix controller and a multiplier of the target tracking device according to the second embodiment of the present invention.

FIG. 6 is a block diagram illustrating a configuration of a target tracking device according to a third embodiment of the present invention.

FIG. 7 is a flowchart showing an operation of the model controller of the target tracking device according to the third embodiment of the present invention.

FIG. 8 is a flowchart showing an operation of the model controller of the target tracking device according to the fourth embodiment of the present invention.

FIG. 9 is a block diagram showing a configuration of a target tracking device according to Embodiment 5 of the present invention.

FIG. 10 is a flowchart showing an operation of a turning time determiner of the target tracking device according to the fifth embodiment of the present invention.

FIG. 11 is a flowchart showing operations of an attenuation gain calculator and a model setting device of the target tracking device according to the fifth embodiment of the present invention.

FIG. 12 is a block diagram illustrating a configuration of a conventional target tracking device.

FIG. 13 is a flowchart showing the operation of a conventional target tracking device.

FIG. 14 is a diagram illustrating N constant acceleration vectors (including a zero acceleration vector) set as a target motion model.

FIG. 15 is a diagram showing a positional relationship between error ellipsoids having different flattening rates around prediction positions of a plurality of motion models.

[Explanation of symbols]

1 target observation device, 2 gate judgment device, 3 reliability calculator, 4 smoother, 5 reliability memory, 6 smoother memory,
7 Predictor, 8 First delay circuit, 9 Predictor for each model, 10 Gain matrix calculator, 11 Model setting device, 1
2 input vector smoother, 13 input vector predictor,
14 Second delay circuit, 15 display device, 16 fourth delay circuit, 17 third delay circuit, 18 error ellipsoid overlap determiner, 19 error covariance matrix controller, 20 multiplier, 21 model controller, 22 turn duration time judgment device, 2
3 Attenuation gain calculator.

 ──────────────────────────────────────────────────続 き Continuing on the front page (72) Inventor Shingo Tsujimichi 2-3-2 Marunouchi, Chiyoda-ku, Tokyo Inside Mitsubishi Electric Corporation (72) Inventor Yoshio Kosuge 2-3-2 Marunouchi, Chiyoda-ku, Tokyo F term in Mitsubishi Electric Corporation (reference) 5J070 AC01 AC06 AD01 AJ13 AK22 BB01 BB16

Claims (11)

[Claims] The use of a Kalman filter for N motion models composed of N constant acceleration vectors makes it possible to extremely increase the reliability of a specific motion model, and to oscillate the phenomenon that the motion model is switched every sampling and the oscillation phenomenon. In a target tracking device equipped with a control device that suppresses the accompanying deterioration of tracking accuracy and the uniformity of reliability of each motion model and the resulting deterioration of observation noise reduction performance, each of the N motion models calculated by the Kalman filter is used. Based on the predicted value of and the prediction error covariance matrix,
An error ellipsoid overlap degree determination means for determining the overlap degree of the error ellipsoid which is an equal probability line indicating the expected existence probability of the target centered on the predicted positions of the N motion models; and the error ellipsoid overlap degree determination means Error ellipsoid enlarging / reducing means for enlarging and reducing the error ellipsoid so that no blank area is generated between the error ellipsoids based on the degree of overlap of the error ellipsoid determined by Characteristic target tracking device. 2. The target tracking device according to claim 1, wherein the error ellipsoid enlarging / reducing means enlarges / reduces the error ellipsoid by controlling the number of target motion models and the magnitude of a constant acceleration vector. A target tracking device characterized by causing the target to be tracked. 3. The use of a Kalman filter for N motion models composed of N constant acceleration vectors, the reliability of a specific motion model becomes extremely high, and the oscillation phenomenon in which the motion model is switched every sampling and the oscillation phenomenon. In the target tracking device equipped with a control device that suppresses the accompanying deterioration of the tracking accuracy and the uniformity of the reliability of each motion model and the resulting deterioration of the observation noise reduction performance, the Kalman filter keeps turning the target during one sampling. The target turn duration time is estimated from the target model information observed by the target observation device and the target maximum acceleration performance based on the above assumption, and the change in speed or acceleration due to the turn is excessive. Turning duration determining means for determining the possibility of calculation, and an external input term in a calculation formula for calculating the reliability And a damping gain calculating means for calculating a damping gain for attenuating the influence of the target. 4. The target tracking device according to claim 1, wherein the error ellipsoid overlap degree determination means includes a major axis, a minor axis, and a median diameter of the error ellipsoid of each motion model, and the respective motions. Determining a degree of overlap of the error ellipsoid by comparing a ratio of a Euclidean distance, which is a distance between the model and a predicted position of the constant velocity motion model, to a preset threshold value. Tracking device. 5. The target tracking device according to claim 1, wherein the error ellipsoid overlap degree determination unit calculates the Euclidean distance by a residual error covariance matrix in a calculation formula for calculating the reliability. A target tracking apparatus, wherein the degree of overlap of the error ellipsoid is determined by comparing the normalized residual quadratic form with a preset threshold value. 6. The target tracking device according to claim 3, wherein the turning continuation time determining means compares the ratio between the determined turning continuation time of the target and the sampling interval with a preset threshold value, thereby turning the target. A target tracking device for determining a possibility that a change in speed or acceleration due to an error is excessively calculated. 7. The target tracking apparatus according to claim 1, wherein a length, a short diameter, and a middle diameter of an error ellipsoid of each motion model and a distance between each motion model and a predicted position of the constant velocity motion model. A target tracking device, further comprising: a multiplication coefficient calculation unit that calculates a multiplication coefficient of an error covariance matrix in a calculation formula for calculating the reliability based on a ratio with the Euclidean distance. 8. The target tracking device according to claim 7, wherein the multiplication coefficient calculation means calculates a distance between an observation value observed by the target observation device and a predicted position of each motion model by a residual error covariance matrix. A target tracking device, characterized in that the error ellipsoid is enlarged and reduced using a normalized residual quadratic form. 9. The target tracking apparatus according to claim 2, wherein the major axis, the minor axis, and the medium axis of the calculated error ellipsoid of the motion model and the distance between each of the motion models and the predicted position of the constant velocity motion model. Using a ratio to the Euclidean distance that is, a constant acceleration vector of each motion model is recalculated, and a model control means is provided for adjusting a predicted position between the motion models in which the error ellipsoid does not overlap with each other. Tracking device. 10. The target tracking apparatus according to claim 2, wherein the major, minor, and medium diameters of the error ellipsoid of the calculated motion model and the distance between each of the motion models and the predicted position of the constant velocity motion model. A target tracking device characterized by comprising a model control means for specifying a region where the error ellipsoids of each motion model do not overlap with each other using the ratio to the Euclidean distance, and adding a new motion model to the region. . 11. The target tracking device according to claim 3, wherein the damping gain calculating means calculates the damping gain of the constant acceleration vector based on the sampling interval and the estimated turning duration of the target. Characteristic target tracking device.