An adaptive new multi-extended target tracking method

Abstract

The invention discloses an adaptive new multi-extended target tracking method. A new measurement-driven adaptive new distribution method is used to handle the tracking problem of new target positions that are unknown. First, the correlation between measurement and target is expressed by likelihood and proximity, and then the new target is generated near the measurement value by using the correlation. Then, the unknown detection probability is described by Beta distribution, and the extended target state is described by Gamma-Gaussian inverse Wishart distribution. Then, the target state and the detection probability are modeled as augmented states, so as to realize the tracking problem of multiple extended targets under unknown detection probability. In addition, in each iteration, the PMBM filter density after truncation and correction is obtained by Gibbs sampler, so as to obtain a positive 1-1 vector with high weight, so as to improve the efficiency of the filter without losing accuracy, and solve the technical problem that the multi-extended target tracking effect is poor when both the detection probability and the new target position are unknown in the prior art.

Description

Technical Field

[0001] The present invention relates to the field of detection and tracking technology, and in particular to an adaptive new multi-extended target tracking method. Background Art

[0002] In the existing multi-target tracking methods, most of the filters based on random finite sets (RFS), such as the standard PMBM conjugate prior, assume that the detection probability and the target new position are estimated from offline training data, that is, they are all known prior information. However, it is worth noting that this knowledge is uncertain in actual situations. Research on multiple extended targets, adaptive target new algorithms, and PMBM conjugate priors under unknown detection probabilities has not yet been achieved. Summary of the invention

[0003] The purpose of the present invention is to provide an adaptive new multi-extended target tracking method, aiming to solve the technical problem in the prior art that the multi-extended target tracking effect is poor when both the detection probability and the new target position are unknown.

[0004] To achieve the above object, the present invention adopts an adaptive new multi-extended target tracking method, comprising the following steps:

[0005] Initialize system parameters and adapt to new expansion targets;

[0006] Combining the unknown detection probabilities to form an augmented vector of the extended target;

[0007] Filtering for prediction and correction;

[0008] A Gibbs sampler is used to truncate the filtered density and extract high-weighted global hypotheses.

[0009] In the process of initializing system parameters and adaptively generating new extended targets, the likelihood and proximity between the target and the measurement are first described by Gaussian distribution and Euclidean metric respectively, and then the likelihood function and proximity are used to represent the correlation between the target and the measurement. Finally, this correlation is used to generate new targets near the measurement.

[0010] Among them, after the new target is generated, the new target is combined with the unknown target to form a new set of unknown targets, and then the Beta distribution is used to describe the unknown detection probability, and the unknown target is combined with the unknown detection probability to form the augmented state of the unknown target. Then, the Gamma-Gaussian inverse Wishart distribution is used to describe the prior target state, and then combined with the unknown detection probability to form the augmented state space, thereby obtaining the augmented state of the prior target.

[0011] After the augmented states of the unknown and prior targets are obtained, they are predicted and updated through filters, which includes recursively filtering the unknown targets and multi-Bernoulli mixture components using Poisson multi-Bernoulli mixture filtering to obtain the posterior information of the Poisson multi-Bernoulli mixture density.

[0012] After the target is predicted and updated, the Gibbs sampler is used to truncate the filter density and extract high-weight global hypotheses. According to the different sources of measurement, a cost matrix is ​​generated, and then the set of corresponding relationships between the target and the measurement is obtained through the Gibbs sampler based on the Markov chain. Finally, the estimated state of the target is extracted.

[0013] The correspondence between the target and the measurement is specifically reflected as an unordered positive 1-1 vector.

[0014] An adaptive new multi-extended target tracking method of the present invention processes the tracking problem of new target positions with unknown locations through a new measurement-driven adaptive new distribution method. First, the correlation between the measurement and the target is expressed by likelihood and proximity, and then the correlation is used to generate new targets near the measurement value. Subsequently, the unknown detection probability is described by Beta distribution, and the extended target state is described by Gamma-Gaussian inverse Wishart distribution. Then, the target state and the detection probability are modeled as augmented states to achieve the tracking problem of multiple extended targets under unknown detection probability. In addition, in each iteration, the PMBM filter density after truncation and correction is obtained by Gibbs sampler to obtain a high-weight positive 1-1 vector, thereby improving the efficiency of the filter without losing accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0016] Figure 1 It is a schematic diagram of the steps of an adaptive new multi-extended target tracking method of the present invention.

[0017] Figure 2 It is a schematic diagram of the execution flow of an adaptive new multi-extended target tracking method of the present invention.

[0018] Figure 3 It is a real target trajectory diagram of a specific embodiment of the present invention.

[0019] Figure 4 It is a measurement distribution diagram corresponding to the real target trajectory of a specific embodiment of the present invention.

[0020] Figure 5 This is a comparison chart of GOSPA distance estimation between the present method and the standard algorithm in embodiment scenario 1.

[0021] Figure 6 This is a comparison chart of OSPA distance and potential estimation between the present method and the standard algorithm in embodiment scenario 1.

[0022] Figure 7 This is a comparison chart of GOSPA distance estimation between the method in embodiment scenario 2 and the standard algorithm.

[0023] Figure 8 This is a comparison chart of OSPA distance estimation, potential estimation and detection probability estimation of the method in embodiment scenario 2 and the standard algorithm.

[0024] Fig. 9 This is a comparison chart of GOSPA distance estimation between the present method and the standard algorithm in embodiment scenario 3.

[0025] Fig.10 This is a comparison chart of OSPA distance estimation, potential estimation and detection probability estimation of the present method and the standard algorithm in embodiment scenario 3.

[0026] Fig.11 This is a comparison chart of GOSPA distance estimation between the method in embodiment scenario 4 and the standard algorithm.

[0027] Fig.12 This is a comparison chart of the OSPA distance estimation and potential estimation of the present method and the standard algorithm in embodiment scenario 4. DETAILED DESCRIPTION

[0028] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and should not be construed as limiting the present invention.

[0029] In this application document, the following terms are used: 'std' represents standard Poisson multi-Bernoulli mixture filtering based on a fixed newborn distribution, 'tb' represents filtering based on an adaptive newborn distribution, 'gbs' represents Gibbs sampling, 'BGGIW' represents filtering based on an unknown detection probability, 'true value' represents the actual but unknown detection probability, and 'PHD' represents probability hypothesis density filtering.

[0030] See also Figure 1 The present invention proposes an adaptive new multi-extended target tracking method, comprising the following steps:

[0031] S1: Initialize system parameters and adapt to new expansion targets;

[0032] S2: Combining the unknown detection probability to form an augmented vector of the extended target;

[0033] S3: filtering for prediction and correction;

[0034] S4: A Gibbs sampler is used to truncate the filtered density and extract high-weighted global hypotheses.

[0035] In the process of initializing system parameters and adaptively generating new extended targets, the likelihood and proximity between the target and the measurement are first described by Gaussian distribution and Euclidean metric respectively, and then the likelihood function and proximity are used to represent the correlation between the target and the measurement. Finally, this correlation is used to generate new targets near the measurement.

[0036] In the process of combining the unknown detection probability to form the augmented vector of the extended target, the unknown detection probability is described by using the Beta distribution, and the target state is described by using the Gamma-Gaussian Inverse Wishart distribution, which are then combined to form the augmented state space.

[0037] The augmented state space is formed by merging the unknown detection probability space and the extended target state space, and the transition density function of the augmented state is obtained by multiplying the transition density function of the target and the unknown detection probability evolution function.

[0038] In the process of filtering prediction and correction, the new targets, potential targets and multi-Bernoulli mixture components are recursively filtered through Poisson multi-Bernoulli mixture filtering to obtain the posterior information of Poisson multi-Bernoulli mixture density.

[0039] Furthermore, in the recursive operation, the new targets, potential targets and multi-Bernoulli mixtures are first predicted, and then the prediction results are corrected, which includes the missed detection of new targets, potential targets and multi-Bernoulli mixture components, and the update of new targets, potential targets and multi-Bernoulli mixture components detected for the first time.

[0040] In the process of using the Gibbs sampler to truncate the filter density and extract high-weight global hypotheses, a cost matrix is ​​first generated according to different sources of measurement, and then a set of unordered positive 1-1 vectors is obtained through a Gibbs sampler based on a Markov chain.

[0041] See also Figure 2 , the following is a detailed description of each step:

[0042] S1. Initialize system parameters and adapt to new expansion targets.

[0043] Specifically, the system parameters are initialized, including: the cutoff parameter c of the reference distance, the penalty parameter p of the outlier, the plane size N of the tracking scene x ×N y, transfer matrix F, process noise Q, survival probability P s Adaptive new targets using likelihood and proximity between measurements and targets where α i and β i is the gamma distribution parameter describing the number of target measurements, m i and P i is the Gaussian distribution parameter describing the target centroid and covariance, v i and V i is the inverse Wishart distribution parameter describing the target shape, s i and t i is the Beta distribution parameter describing the detection probability, λ exp is the mean number of new targets.

[0044] Assume that the measurement received at the previous moment (here assumed to be the kth moment) is z k,1 =[x k,1 ,y k,1 ] T ,…,z k,N =[x k,N ,y k,N ] T ∈Z, N is the number of measurements. Assume that in the calibrated multi-Bernoulli mixture (MBM), the center of mass of the target is Calculate the likelihood between the measurement and the target using a Gaussian distribution

[0045]

[0046] Where Σ represents the covariance corresponding to the target, det(·) represents the determinant of the matrix, and then the likelihood truncation threshold λ is used g Select the corresponding measurement set Z g . Then the target centroid is calculated by Euclidean metric and measurement k,g =[x k,g ,y k,g ]∈Z g Proximity

[0047]

[0048] Reusing the proximity cutoff threshold λ d Select the corresponding measurement set Z d .

[0049] Since the prior information of the target is unknown, each measurement z k,d ∈Z d It is possible to generate new targets, that is, each measurement will generate a corresponding BGGIW component, and its weight is

[0050]

[0051] where w max represents the maximum weight of the new BGGIW component, λ exp is the expected value of the number of new BGGIW components.

[0052] S2. Combining the unknown detection probability to form an augmented vector of the extended target, specifically using Beta distribution to describe the unknown detection probability, using Gamma-Gaussian inverse Wishart distribution to describe the target state, and then combining them into an augmented vector.

[0053] Furthermore, let X be the extended target state space, X (^) represents the unknown detection probability space, then the augmented state space is expressed as

[0054]

[0055] Where × represents the Cartesian product, each augmented state By extending the target state x∈X and increasing the detection probability η∈X (^) =[0,1]. The transfer density function can be expressed as

[0056]

[0057] where f k|k-1 (·|·) represents the target state transition density, is the evolution function of the detection probability. The survival probability and detection probability of the augmented state are

[0058]

[0059]

[0060] S3. Filtering is performed for prediction and correction, specifically filtering the new targets, potential targets and Poisson Multi-Bernoulli mixtures respectively to obtain the posterior information of the Poisson Multi-Bernoulli mixture density.

[0061] Furthermore, the strength of the new target and potential target at time k-1 is expressed by BGGIW hybrid, which are

[0062]

[0063]

[0064] Where B(·) is the beta distribution, which represents the change of unknown detection probability, and s and t are its corresponding parameters. GGIW(·) is the GGIW distribution, which consists of the gamma probability density function, the multivariate Gaussian probability density function and the inverse Chatter distribution. The gamma probability density function is used to describe the change in the number of measurements generated by the extended target, the multivariate Gaussian probability density function is used to describe the change in the center of mass of the extended target and its corresponding covariance, and the inverse Chatter distribution is used to represent the change in the shape of the extended target, and α, β, m, P, v, V are its corresponding parameters. Let the multi-Bernoulli mixture density at time k-1 be expressed as

[0065]

[0066]

[0067] in represents the probability density function of the nth Bernoulli in the sth multi-Bernoulli, The corresponding existence probability is, represents the weight of the sth multi-Bernoulli, Π s Represents the index set of the Bernoulli in the s-th multi-Bernoulli.

[0068] First, the predictions for new targets, potential targets, and multi-Bernoulli mixtures are made. The prediction results for new targets and potential targets are also in the mixed form of BGGIW. The prediction results for each Bernoulli in the multi-Bernoulli mixture are in a single BGGIW form, so the overall prediction can also be represented by the mixed form of BGGIW. The second step is to calibrate the prediction results. The calibration stage is divided into the following four steps:

[0069] 1) Missed detection of new and potential targets.

[0070] 2) New and potential targets are detected for the first time.

[0071] 3) Missed detection of Poisson-Multi-Bernoulli mixtures.

[0072] 4) Correction for Poisson-Multi-Bernoulli mixture.

[0073] Likewise, the composition after the correction stage can be represented by a mixed form of BGGIW.

[0074] S4, using Gibbs sampler to truncate the filtered density and extract high-weighted global hypotheses.

[0075] Specifically, based on the different sources of measurement, a cost matrix is ​​generated

[0076]

[0077]

[0078] C miss,i =-ln(1-r k,i +r k,i (1-P d ))

[0079] where r k,m and r k,i is the probability of existence of the surviving Bernoulli component, is its corresponding probability density function, l(·) is the prediction likelihood function, W n is the measurement cell obtained by measurement division,<a,b> represents the integral of a and b, is the probability density function of the potential target, κ C is the clutter intensity, P d is the detection probability.

[0080] After obtaining the cost matrix, the global hypothesis of data association is obtained through the Gibbs sampler based on Markov chain. The transfer kernel function of the Gibbs sampler is expressed as

[0081]

[0082] where υ' represents a global assumption about the joint allocation problem, and its physical meaning is the detection of n1 surviving Bernoulli components. represents an unordered set of positive 1-1 vectors, which represent the association between the measurement cell and the target, and It represents a positive 1-1 vector.

[0083] The present invention also verifies the adaptive new multi-extended target tracking method through simulation experiments:

[0084] 1. Simulation conditions: The present invention is simulated on a computer with an Intel(R) Core(TM) i7-7700 CPU@3.60GHz and 8.0GB memory processor using MATLAB R2020a software.

[0085] 2. Simulation scene setting: Consider a two-dimensional scene of [-200,200]m×[-200,200]m, where a total of 27 randomly generated targets move 100 time steps in the surveillance area through a uniform motion model, such as Figure 3 As shown, the corresponding measurement distribution is Figure 4 The Poisson rate λ is given in x =20, detection probability P d = 0.9, the number of measurements generated by each target follows the mean λ x = 8 Poisson distribution. The transfer matrix F and process noise Q are

[0086]

[0087] Where T s =1 represents the time interval, σ=0.1 represents the noise driving parameter.

[0088] Fixed freshman distribution D b The intensity of (x) is expressed in the form of a GGIW mixture

[0089]

[0090] w b,1 =w b,2 =0.02,w b,3 =w b,4 =0.03

[0091] m b,1 =[-75,-75,0,0],m b,2 =[-75,75,0,0]

[0092] m b,3 =[75,75,0,0],m b,4 =[75,-75,0,0]

[0093] α b,i =32,β b,i =4,v b,i =12

[0094] P b,i =diag([1,1,1,1] T ) 2 ,V b,i =diag([12,12] T ), where i∈{1,2,3,4}.

[0095] In scenario 1, Figure 5 and Figure 6As shown, the cardinality estimates for the adaptive and fixed birth distributions are almost equal except for a few time steps after the birth of the target. In addition, after the birth of the new target, the cardinality estimate for the fixed birth distribution is larger than that for the adaptive birth distribution because the expected number of target births for the adaptive birth distribution is assigned to more BGGIW components. Therefore, the new target has a lower weight and the cardinality estimate will grow at a slower rate. In the time steps before and after the birth of the target, the OSPA distance and GOSPA distance (including LE, ME and FE) of the adaptive birth distribution are larger than those of the fixed birth distribution, while at other times, the two are almost equal. For the adaptive birth distribution, all estimates start from the second time step because the measurements of the first time step are used to generate the new target. The estimated averages are given in Table 1. It can be seen that the use of the Gibbs sampler has no obvious effect on the accuracy of the estimation, but can greatly improve the efficiency of the estimation.

[0096] Table 1: Estimated mean values

[0097] filter std PMBM-tb PMBM-tb-gbs OSPA 0.918 1.150 1.171 GOSPA 7.506 8.744 8.884 LE 5.631 5.844 6.124 ME 0.229 0.362 0.335 FE 0.146 0.218 0.217 Time 245.827 775.073 413.713

[0098] In scenario 2, Figure 7 As shown in Figure 2, the OSPA distance and potential estimates of the robust PMBM filter are very close to those of the standard PMBM filter. Figure 8 The comparison of GOSPA distance also reflects the similar conclusion as OSPA distance. Since the robust PMBM filter has the ability of online adaptive adjustment, it is slightly more robust to ME than the standard PMBM filter. As can be seen from Table 2, as expected, the Gibbs sampler can greatly improve the estimation efficiency without changing the estimation accuracy.

[0099] Table 2: Estimated mean values

[0100] filter std BGGIW-PMBM BGGIW-PMBM-gbs OSPA 0.918 0.942 0.991 GOSPA 7.506 7.657 7.989 LE 5.631 5.772 5.959 ME 0.229 0.077 0.099 FE 0.146 0.301 0.307 Time 245.827 1022.45 564.494

[0101] In scenario 3, Fig. 9 and Fig.10 As shown, similar to the previous conclusion, except for the time near the birth of the new target, the estimated value of the filter using the adaptive birth distribution is very close to the estimated value of the filter using the fixed birth distribution. The estimated value of the unknown probability filter is also close to the estimated value of the known probability filter. The estimated average values ​​are given in Table 3. It can be seen that the Gibbs sampling method greatly reduces the calculation time of the filter.

[0102] Table 3: Estimated mean values

[0103] filter std BGGIW-PMBM BGGIW-PMBM-tb BGGIW-PMBM-tb-gbs PMBM-tb OSPA 0.821 0.936 1.108 1.162 1.118 GOSPA 6.513 7.217 8.744 8.752 8.075 LE 5.533 5.472 5.659 5.707 5.410 ME 0.084 0.212 0.312 0.324 0.395 FE 0.112 0.317 0.305 0.285 0.138 Time 217.44 248.035 1119.091 618.319 389.277

[0104] In scenario 4, Fig.11 and Fig.12 As shown in the figure, it can be seen that the performance of the BGGIW-PHD-tb filter is much worse than that of the other filters, which can be well explained by the poor tolerance of the PHD filter when the number of targets is large. The BGGIW-PMBM-tb filter has the best performance in all aspects.

[0105] Aiming at the problem of multi-extended target tracking with unknown detection probability and unknown birth position, the present invention proposes a BGGIW-PMBM algorithm, in which the unknown detection probability is described by Beta distribution and the target birth intensity is modeled by adaptive rebirth distribution. The Gibbs sampler is used to reduce the computational burden caused by data association. The superiority of the proposed algorithm is verified through four scenarios, in which scenario 1 proves the efficiency of the filter under Gibbs sampling and the stability of the adaptive rebirth of the target; scenario 2 proves the effectiveness of the filter when the unknown detection probability is described by Beta distribution; scenario 3 gives the efficiency of the filter under the combination of three methods: adaptive rebirth algorithm, unknown detection probability algorithm and Gibbs sampling; and scenario 4 makes a comprehensive comparison between the PMBM filter and other filters, thereby proving the accuracy of the PMBM algorithm.

[0106] What is disclosed above is only a preferred embodiment of the present invention, and it certainly cannot be used to limit the scope of rights of the present invention. Ordinary technicians in this field can understand that all or part of the processes of the above embodiment and equivalent changes made according to the claims of the present invention still fall within the scope of the invention.

Claims

1. An adaptive new multi-extended target tracking method, It is characterized in that The following steps are involved: Initialize system parameters and adapt to new expansion targets; Combining the unknown detection probabilities to form an augmented vector of the extended target; The process of combining the unknown detection probability to form the augmented vector of the extended target is specifically, after generating the new target, combining the new target with the unknown target to form a new unknown target set, then using the beta distribution to describe the unknown detection probability, combining the unknown target with the unknown detection probability to form the augmented state of the unknown target, and then using the gamma-Gaussian inverse Wishart distribution to describe the prior target state, and then combining the unknown detection probability to form the augmented state space, so as to obtain the augmented state of the prior target; Filtering for prediction and correction; Specifically, the new targets, potential targets and Poisson multi-Bernoulli mixtures are filtered respectively to obtain the posterior information of the Poisson multi-Bernoulli mixture density; A Gibbs sampler is used to truncate the filter density and extract high-weighted global hypotheses; Among them, the transfer kernel function of the Gibbs sampler is expressed as where v' represents a global assumption about the joint allocation problem, and its physical meaning is the detection of n1 surviving Bernoulli components.

2. The adaptive new multi-extended target tracking method according to claim 1, It is characterized in that In the process of initializing system parameters and adaptively generating new extended targets, the likelihood and proximity between the target and the measurement are first described by Gaussian distribution and Euclidean metric respectively, and then the likelihood function and proximity are used to represent the correlation between the target and the measurement. Finally, this correlation is used to generate new targets near the measurement.

3. The adaptive new multi-extended target tracking method as claimed in claim 2, It is characterized in that After obtaining the augmented states of the unknown and prior targets, they are predicted and updated through filters, which includes recursively filtering the unknown targets and multi-Bernoulli mixture components using Poisson multi-Bernoulli mixture filtering to obtain the posterior information of the Poisson multi-Bernoulli mixture density.

4. The adaptive new multi-extended target tracking method as claimed in claim 3, It is characterized in that After the target is predicted and updated, the Gibbs sampler is used to truncate the filter density and extract high-weight global hypotheses. According to the different sources of measurement, a cost matrix is ​​generated, and then the set of correspondences between the target and the measurement is obtained through the Gibbs sampler based on the Markov chain. Finally, the estimated state of the target is extracted.

Images