Multi-sensor multi-target tracking method based on a posteriori track estimation
By constructing an association cost matrix and track matching technique based on posterior track estimation, the problems of excessive communication load and erroneous tracks in multi-sensor multi-target tracking are solved, achieving high-precision, robust and real-time multi-sensor multi-target tracking.
Patent Information
- Application Number
- CN202210395505.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-04-15
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2042-04-15
AI Technical Summary
Existing multi-sensor multi-target tracking technologies suffer from problems such as excessive communication load, easy generation of counterintuitive erroneous tracks, and target miss detection. In particular, when extending the track probability hypothesis density filtering method to the field of multi-sensor multi-target tracking based on the GCI criterion, it is difficult to achieve effective track confirmation and information fusion.
A posterior trajectory estimation-based approach is adopted. By constructing an association cost matrix and using trajectory matching technology, erroneous trajectories are eliminated, and the trajectories are directly output. The complementary information characteristics of multiple sensors are used for trajectory fusion, which reduces computational complexity and communication load.
It improves the accuracy and robustness of trajectory estimation, reduces the probability of erroneous trajectories, enhances the probability and real-time performance of target detection, and achieves efficient multi-sensor multi-target tracking.
Smart Images

Figure CN114757971B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of target tracking, and in particular relates to a multi-sensor multi-target tracking method based on posterior track estimation. Background Art
[0002] Distributed multi-sensor multi-target tracking (DMMT) refers to the comprehensive use of noisy observation information provided by multiple sensors, which, after appropriate fusion processing, can eliminate the false and retain the true, and make up for the shortcomings of the others, so as to obtain a more comprehensive and accurate target track status than a single sensor. It is a key technology of dual-use military and civilian reconnaissance and surveillance systems, and mainly faces two major challenges: multi-target tracking and distributed information fusion.
[0003] The most typical multi-target tracking methods include three categories: multi-hypothesis tracking, joint probabilistic data association, and random finite sets. From a Bayesian perspective, the finite set statistics theory based on random finite sets is an ideal framework for solving the multi-target tracking problem. Within this framework, probability hypothesis density, multi-target multi-Bernoulli, and generalized label multi-Bernoulli filters have received widespread attention from a large number of scholars. However, the first two can only provide instant multi-target motion state estimation, and cannot confirm the multi-target identity information estimated at the previous and next moments, that is, it is impossible to form a track, while the generalized label multi-Bernoulli filter involves the prediction and update of a large number of multi-target indices and has a high computational complexity.
[0004] In recent years, Garcia-Fernandez et al. proposed a trajectory probability hypothesis density (TPHD) filtering method, which performs well in terms of estimation accuracy and computational complexity, provides accurate a posteriori track estimation, and can also improve the estimation accuracy of the track state at past moments.
[0005] Distributed information fusion can make up for the shortcomings of limited perception capabilities of a single node, but due to the existence of relevant information between sensors that is difficult to quantitatively characterize, the fusion method based on Bayesian optimality is difficult to implement and has poor robustness. In recent years, the generalized covariance inverse (GCI) fusion criterion has been successfully applied to various multi-target trackers, such as probability hypothesis density filters. Although these distributed information fusion algorithms are rigorous in theory, they still cannot confirm the identity information of multiple targets estimated at the previous and next moments, that is, they cannot form a track, and there are problems such as delay in detecting new targets and high computational complexity. GCI fusion based on label multi-Bernoulli filters can estimate tracks, but there is a problem of inconsistent labels (different nodes assign different labels to the same target). Even if solutions such as label-free and label matching are adopted, there are still challenges such as computational difficulties and local optimality.
[0006] Given the excellent performance of the track probability hypothesis density (TPHD) filtering method in multi-target state estimation and track management, it is of great application value to extend it to the field of multi-sensor multi-target tracking based on the GCI criterion. However, direct combined application will have the following three major problems: first, the communication load required to transmit the posterior track probability hypothesis density between nodes is too large; second, simple GCI fusion is prone to produce counter-intuitive erroneous tracks; third, there is a common phenomenon of missed targets due to inconsistent node potential distribution. Summary of the invention
[0007] In view of the multi-sensor multi-target tracking problem existing in the prior art, and considering that simply extending the track probability hypothesis density (TPHD) filtering method to the field of multi-sensor multi-target tracking based on the GCI criterion will face challenges such as excessive communication load, easy generation of counter-intuitive erroneous tracks and target missed detection, the present invention provides a multi-sensor multi-target tracking method based on a posteriori track estimation to enhance the real-time, accuracy and robustness of multi-sensor multi-target tracking.
[0008] In order to achieve the above technical objectives, the technical solution proposed by the present invention is:
[0009] The multi-sensor multi-target tracking method based on a posteriori track estimation includes:
[0010] Based on the prior track probability hypothesis density of each sensor, the posterior track probability hypothesis density of each sensor is obtained by using the track probability hypothesis density filtering method, and the posterior track estimation set corresponding to the posterior track probability hypothesis density of each sensor is obtained;
[0011] Constructing a correlation cost matrix between different sensors, and dividing the posterior tracks in the posterior track estimation sets of different sensors into correlated tracks and non-correlated tracks based on the correlation cost matrix between different sensors;
[0012] The probability hypothesis density of the associated tracks between different sensors is fused to obtain the fused probability hypothesis density of the associated tracks between different sensors;
[0013] Eliminate the erroneous tracks among the non-correlated tracks between different sensors, fuse the probability hypothesis density of the non-correlated tracks after eliminating the erroneous tracks with the probability hypothesis density of the fused correlated tracks, and obtain the probability hypothesis density of the fused tracks between different sensors;
[0014] According to the fusion track probability hypothesis density among different sensors, the multi-sensor estimated track state is extracted.
[0015] As a preferred embodiment of the present invention, constructing the association cost matrix between different sensors includes the following steps:
[0016] According to the posterior track probability hypothesis density of different sensors, the similarity matrix of the posterior track estimation sets of different sensors is calculated;
[0017] The association matrix of the posterior track estimation set corresponding to different sensors is constructed, the cost of the association matrix is set based on the similarity matrix, and the association matrix with the minimum cost is sought using the optimization algorithm as the final association cost matrix.
[0018] As a preferred embodiment of the present invention, the method for eliminating erroneous tracks is: by setting a threshold, eliminating erroneous tracks whose lengths of non-correlated tracks are less than the set threshold.
[0019] On the other hand, the present invention provides a multi-sensor multi-target tracking device based on a posteriori track estimation, comprising:
[0020] The first module is used to obtain the posterior track probability hypothesis density of each sensor based on the prior track probability hypothesis density of each sensor by using the track probability hypothesis density filtering method, and obtain the posterior track estimation set corresponding to the posterior track probability hypothesis density of different sensors;
[0021] The second module is used to construct an association cost matrix between different sensors, and divide the posterior tracks in the posterior track estimation sets of different sensors into associated tracks and non-associated tracks based on the association cost matrix between different sensors;
[0022] The third module is used to fuse the probability hypothesis density of the associated tracks between different sensors to obtain the fused associated track probability hypothesis density between different sensors;
[0023] The fourth module is used to eliminate the erroneous tracks in the non-correlated tracks between different sensors, and fuse the probability hypothesis density of the non-correlated tracks after eliminating the erroneous tracks with the probability hypothesis density of the fused correlated tracks to obtain the probability hypothesis density of the fused tracks between different sensors;
[0024] The fifth module is used to extract the multi-sensor estimated track state according to the fusion track probability hypothesis density between different sensors.
[0025] On the other hand, the present invention provides a computer-readable storage medium having a computer program stored thereon, characterized in that when a processor executes the computer program, the following steps are implemented:
[0026] Based on the prior track probability hypothesis density of each sensor, the posterior track probability hypothesis density of each sensor is obtained by using the track probability hypothesis density filtering method, and the posterior track estimation set corresponding to the posterior track probability hypothesis density of each sensor is obtained;
[0027] Constructing a correlation cost matrix between different sensors, and dividing the posterior tracks in the posterior track estimation sets of different sensors into correlated tracks and non-correlated tracks based on the correlation cost matrix between different sensors;
[0028] The probability hypothesis density of the associated tracks between different sensors is fused to obtain the fused probability hypothesis density of the associated tracks between different sensors;
[0029] Eliminate the erroneous tracks among the non-correlated tracks between different sensors, fuse the probability hypothesis density of the non-correlated tracks after eliminating the erroneous tracks with the probability hypothesis density of the fused correlated tracks, and obtain the probability hypothesis density of the fused tracks between different sensors;
[0030] According to the fusion track probability hypothesis density among different sensors, the multi-sensor estimated track state is extracted.
[0031] In another aspect, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:
[0032] Based on the prior track probability hypothesis density of each sensor, the posterior track probability hypothesis density of each sensor is obtained by using the track probability hypothesis density filtering method, and the posterior track estimation set corresponding to the posterior track probability hypothesis density of each sensor is obtained;
[0033] Constructing a correlation cost matrix between different sensors, and dividing the posterior tracks in the posterior track estimation sets of different sensors into correlated tracks and non-correlated tracks based on the correlation cost matrix between different sensors;
[0034] The probability hypothesis density of the associated tracks between different sensors is fused to obtain the fused probability hypothesis density of the associated tracks between different sensors;
[0035] Eliminate the erroneous tracks among the non-correlated tracks between different sensors, fuse the probability hypothesis density of the non-correlated tracks after eliminating the erroneous tracks with the probability hypothesis density of the fused correlated tracks, and obtain the probability hypothesis density of the fused tracks between different sensors;
[0036] According to the fusion track probability hypothesis density among different sensors, the multi-sensor estimated track state is extracted.
[0037] The present invention uses a strategy of a posteriori track estimation rather than a posteriori track probability hypothesis density, which significantly reduces the communication load. The method also uses track optimal matching technology to reduce the problem of high error rate of fused tracks, and establishes a robust track fusion criterion to solve the problem of missed detection in GCI fusion. Compared with the prior art, the advantages of the present invention are:
[0038] First, the track can be directly output. Compared with the traditional label-assisted DMMT method, the present invention directly estimates and fuses the track based on the track set during the filtering and fusion process, avoiding the complexity and ambiguity caused by introducing labels to construct the track.
[0039] Second, the track estimation accuracy is high. The single-sensor TPHD filter considers multi-step measurement information and has the function of smoothing the track, which improves the track estimation accuracy.
[0040] Third, it is robust. By estimating the track similarity measurement and optimal matching of different sensor nodes, the track fusion of different nodes is only performed under the premise of the same target. The retention of non-correlated tracks fully utilizes the information complementarity of multiple sensors and improves the probability of target detection. By eliminating erroneous tracks, the probability of erroneous tracks is greatly reduced.
[0041] Fourth, the real-time performance is good. Compared with the DMMT method that integrates the posterior probability hypothesis density, the present invention integrates the posterior track estimation, which greatly reduces the computational load, and adopts the track matching technology to make the track fusion completely parallel, further reducing the computational complexity. BRIEF DESCRIPTION OF THE DRAWINGS
[0042] 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 the structures shown in these drawings without paying creative work.
[0043] Figure 1 is a flow chart of an embodiment of the present invention;
[0044] Figure 2 is a flow chart of a specific embodiment of the present invention;
[0045] Figure 3 A schematic diagram of a real target track used in a simulation example of the present invention;
[0046] Figure 4 It is a schematic diagram of the track estimation of the sensor 1 in a simulation example of the present invention;
[0047] Figure 5 It is a schematic diagram of the track estimation of the sensor 2 in a simulation example of the present invention;
[0048] Figure 6 This is a fusion track estimation effect diagram obtained by using the method of the present invention in a simulation example of the present invention;
[0049] Figure 7 This is a comparison diagram of GOSPA metric errors obtained by fusing four methods, namely TPHD-TA-GCI, PHD-GCI, PHD-WAA, and B2B-AA-Flooding, in a simulation example of the present invention;
[0050] Figure 8 This is a comparison diagram of estimated position errors obtained by fusing four methods, namely TPHD-TA-GCI, PHD-GCI, PHD-WAA and B2B-AA-Flooding, in a simulation example of the present invention;
[0051] Fig. 9 This is a comparison diagram of missed detection errors obtained by fusing four methods, namely TPHD-TA-GCI, PHD-GCI, PHD-WAA and B2B-AA-Flooding, in a simulation example of the present invention;
[0052] Fig.10 This is a comparison diagram of the error target errors obtained by fusing the four methods of TPHD-TA-GCI, PHD-GCI, PHD-WAA, and B2B-AA-Flooding in a simulation example of the present invention;
[0053] Fig.11 This is a comparison chart of target number estimation results obtained by integrating four methods, namely TPHD-TA-GCI, PHD-GCI, PHD-WAA and B2B-AA-Flooding, in a simulation example of the present invention;
[0054] Fig.12 It is a structural schematic diagram of an embodiment of the present invention. DETAILED DESCRIPTION
[0055] In order to make the purpose, technical scheme and advantages of the embodiments of the present invention more clearly understood, the following will be used with drawings and detailed descriptions to clearly illustrate the spirit of the contents disclosed by the present invention. After understanding the embodiments of the contents of the present invention, any person skilled in the art can change and modify the techniques taught by the contents of the present invention without departing from the spirit and scope of the contents of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention, but are not intended to limit the present invention.
[0056] In one embodiment, a multi-sensor multi-target tracking method based on a posteriori track estimation is provided, comprising the following steps:
[0057] S1. Based on the prior track probability hypothesis density of each sensor, the posterior track probability hypothesis density of each sensor is obtained by using the track probability hypothesis density filtering method, and the posterior track estimation set corresponding to the posterior track probability hypothesis density of each sensor is obtained;
[0058] S2. constructing an association cost matrix between different sensors, and dividing the posterior tracks in the posterior track estimation sets of different sensors into associated tracks and non-associated tracks based on the association cost matrix between different sensors;
[0059] S3. Fusing the probability hypothesis density of the associated tracks between different sensors to obtain the fused probability hypothesis density of the associated tracks between different sensors;
[0060] S4. Eliminate the erroneous tracks among the non-correlated tracks between different sensors, and fuse the probability hypothesis density of the non-correlated tracks after eliminating the erroneous tracks with the probability hypothesis density of the fused correlated tracks to obtain the probability hypothesis density of the fused tracks between different sensors;
[0061] S5. According to the fusion track probability hypothesis density between different sensors, the multi-sensor estimated track state is extracted.
[0062] In S1, the track probability hypothesis density filtering method is a well-known method in the art. Those skilled in the art can obtain the posterior track probability hypothesis density of each sensor based on the prior track probability hypothesis density of each sensor by using the track probability hypothesis density filtering method, and obtain the posterior track estimation set corresponding to the posterior track probability hypothesis density of each sensor.
[0063] In S2 of one embodiment, constructing an association cost matrix between different sensors includes the following steps:
[0064] (S2.1) calculating a similarity matrix of a posterior track estimation set of different sensors based on the posterior track probability hypothesis density of different sensors;
[0065] (S2.2) constructing a correlation matrix of a posterior track estimation set corresponding to different sensors;
[0066] (S2.3) setting the cost of the association matrix based on the similarity matrix;
[0067] (S2.4) Use the Hungarian algorithm to find the association matrix with the minimum cost as the final association cost matrix.
[0068] In one embodiment, there are γ sensors in total, and the sensor numbers γ∈{γ1,γ2,...,γ Υ}, set the initial time k = 0, initialize the track probability hypothesis density D γ,0 (x), initialize the probability hypothesis density D γ,0 (x) is:
[0069]
[0070] in represents the Gaussian component corresponding to index number i at time k, whose sensor number is γ and weight is Start time is The mean is and the covariance is Index number i∈{1,2,...,I k}, where I k represents the number of Gaussian components at time k. In the symbols involved below, the definition is as follows: (·) k represents the value at time k, then (·)0 represents the initial value, (·) k|k-1 represents the predicted value at time k-1 for time k, (·) (i) represents the physical quantity related to the Gaussian component corresponding to index number i, (·) γ Represents the physical quantity related to sensor γ.
[0071] Reference Figure 2 The multi-sensor multi-target tracking method based on posterior track estimation in this embodiment includes the following steps:
[0072] S1. Based on the prior track probability hypothesis density of each sensor, the posterior track probability hypothesis density of each sensor is obtained by using the track probability hypothesis density filtering method, and the posterior track estimation set corresponding to the posterior track probability hypothesis density of each sensor is obtained.
[0073] (S1.1) Based on the prior track probability hypothesis density D of the sensor at the last moment (k-1 moment) γ,k-1 (X), predict the track probability hypothesis density D of the sensor at the current time (k) γ,k|k-1 (X).
[0074] (S1.2) Based on the track probability hypothesis density D of the sensor at the current moment γ,k|k-1(X) and measurement set in is the measured value, To measure the quantity, obtain the updated sensor track probability hypothesis density D γ,k (x).
[0075] (S1.3) Update the sensor track probability hypothesis density D γ,k The Gaussian component of (x) is pruned to update the track probability hypothesis density D γ,k The low-weight Gaussian component in (x) obtains the sensor's posterior track probability hypothesis density And the corresponding posterior track estimate set, denoted as
[0076] Steps (S1.1) to (S1.3) are performed on each sensor in this embodiment to obtain a posterior track estimation set corresponding to the posterior track probability hypothesis density of each sensor.
[0077] Next, this embodiment uses the posterior track probability hypothesis density of each sensor to fuse and obtain the fused track probability hypothesis density between all sensors. First, the posterior track probability hypothesis density of the first sensor and the second sensor are fused to obtain the first fused track probability hypothesis density The first fused track probability hypothesis density and the posterior track probability hypothesis density of the third sensor are fused to obtain the second fused track probability hypothesis density This process is repeated until all the posterior track probability hypothesis densities of all sensors are fused to obtain the final fused track probability hypothesis density.
[0078] S2 constructs an association cost matrix between sensors γ1 and γ2, and divides the posterior tracks in the posterior track estimation set of sensors γ1 and γ2 into associated tracks and non-associated tracks based on the association cost matrix between sensors γ1 and γ2;
[0079] (S2.1) For sensors γ1 and γ2, the similarity matrix of the posterior track estimation set of sensors γ1 and γ2 at time k is as follows:
[0080]
[0081] Similarity Matrix The element c at row i and column j in i,j represents the similarity between the i-th posterior track in the posterior track estimation set of sensor γ1 and the j-th posterior track in the posterior track estimation set of sensor γ2, and The numbers of Gaussian components representing the posterior track probability hypothesis density of sensors γ1 and γ2 respectively;
[0082] c i,j Calculated by the following formula:
[0083]
[0084] in represents the multi-factor track similarity, v represents the set multi-factor weights. Without loss of generality, in this embodiment, v is set to [0.55, 0.35, 0.10]. ij represents the fuzzy factor set, u ij =[u1,u2,u3] and calculated by the following formula:
[0085]
[0086] in, Represents the Gaussian component mean of the probability hypothesis density of the posterior track of sensor γ1 at time k, that is, the position of the i-th posterior track in the posterior track estimation set of sensor γ1 at time k in the plane rectangular coordinate system on the X axis and its corresponding speed Y-axis position and its corresponding speed Represents the mean of the Gaussian components of the probability hypothesis density of the posterior track of sensor γ2, that is, the position of the jth posterior track in the posterior track estimation set of sensor γ2 at time k in the plane rectangular coordinate system on the X axis and its corresponding speed Y-axis position and its corresponding speed
[0087] (S2.2) Construct the association matrix of the posterior track estimation set corresponding to sensors γ1 and γ2 as follows:
[0088]
[0089] The element in row i and column j of the incidence matrix is a i,j , a i,j is equal to 0 or 1, There is at most one a in each row and column. i,j is equal to 1;
[0090] (S2.3) Set up the correlation matrix The corresponding cost is
[0091] (S2.4) The Hungarian algorithm is used to find the association matrix with the minimum cost as the final association cost matrix between sensors γ1 and γ2.
[0092] (S2.5) Based on the association cost matrix between sensors γ1 and γ2, the posterior tracks in the posterior track estimation set of sensors γ1 and γ2 are divided into associated tracks and non-associated tracks. Specifically, the method is as follows:
[0093] For the i-th posterior track in the posterior track estimation set of sensor γ1, If a i,j =1, and Then it means that the i-th posterior track in the posterior track estimation set of sensor γ1 is associated with the j-th posterior track in the posterior track estimation set of sensor γ2, and the i-th posterior track in the posterior track estimation set of sensor γ1 and the j-th posterior track in the posterior track estimation set of sensor γ2 are associated tracks;
[0094] For the i-th posterior track in the posterior track estimation set of sensor γ1, If a i,j =1, and This means that the i-th posterior track in the posterior track estimation set of sensor γ1 is a non-correlated track;
[0095] For the i-th posterior track in the posterior track estimation set of sensor γ1, If a i,j =1, and This means that the jth posterior track in the posterior track estimation set of sensor γ2 is a non-correlated track.
[0096] S3. Fusing the probability hypothesis density of the associated track between sensors γ1 and γ2 to obtain the fused associated track probability hypothesis density between sensors γ1 and γ2;
[0097] S4. Eliminate the erroneous tracks in the non-correlated tracks between sensors γ1 and γ2, fuse the probability hypothesis density of the non-correlated tracks after eliminating the erroneous tracks with the probability hypothesis density of the fused correlated tracks, and obtain the first fused track probability hypothesis density between sensors γ1 and γ2;
[0098] S5. The first fused track probability hypothesis density between sensors γ1 and γ2 and the posterior track probability hypothesis density of the third sensor are fused using the same method as in steps (S2) to (S4) to obtain a second fused track probability hypothesis density This process is repeated until all the posterior track probability hypothesis densities of all sensors are fused to obtain the final fused track probability hypothesis density. Based on the final fused track probability hypothesis density The multi-sensor estimated track state is extracted.
[0099] In (S1.1), based on the prior track probability hypothesis density D of the sensor at the last moment (k-1 moment) γ,k-1 (X), predict the track probability hypothesis density D of the sensor at the current time (k) γ,k|k-1 (X), comprising the following steps:
[0100] Assume that the prior track probability hypothesis density D at time k-1 is γ,k-1 (X) is:
[0101]
[0102] k The predicted track probability hypothesis density at time is:
[0103] D γ,k|k-1 (X) = D γ,S,k|k-1 (X)+D γ,B,k|k-1 (X)
[0104] Where D γ,S,k|k-1 (X) represents the track probability hypothesis density of the surviving target, D γ,β,k|k-1 (X) represents the track probability hypothesis density of the new target;
[0105] D γ,S,k|k-1 (X) is given by:
[0106]
[0107] in and They are respectively the starting time of the survival track, the mean and covariance of the Gaussian component, and the calculation formula is as follows:
[0108]
[0109]
[0110]
[0111]
[0112] Where n is the single target state dimension, Representation Matrix Columns 1 to 2 in The elements of the column are 0, represents Kronecker product, F and Q represent the target state transfer matrix and process noise covariance matrix respectively, and T represents matrix transpose;
[0113] D γ,B,k|k-1 (x) is given by:
[0114]
[0115] in and are the weights, start time, mean and covariance of the Gaussian components of the new track, respectively. γ,B,k|k-1 is the number of Gaussian components of the new target, where the track start time And i∈{I γ,k-1 +1,...,I γ,k-1 +I γ,B,k|k-1};
[0116] The predicted track probability hypothesis density D at time k γ,k|k-1 (X) is:
[0117]
[0118] Among them I γ,k|k-1 =I γ,k-1 +I γ,B,k|k-1 .
[0119] In (S1.2), based on the track probability hypothesis density D of the sensor at the current moment γ,k|k-1 (X) and measurement set in is the measured value, To measure the quantity, obtain the updated sensor track probability hypothesis density D γ,k (x).
[0120] Update the sensor track probability hypothesis density D γ,k (x) is:
[0121] D γ,k (x) = D γ,E,k (x)+D γ,U,k (x)
[0122] Where D γ,E,k (x) and D γ,U,k (x) represents the probability hypothesis density of missed detection track and the probability hypothesis density of measurement update, respectively:
[0123]
[0124] in represents the detection probability;
[0125]
[0126] is the updated Gaussian component weight, is the track start time, is the updated Gaussian component mean, is the updated covariance, where the weight Calculated by the following formula:
[0127]
[0128] where k(·) represents the clutter intensity function, Calculated by the following formula
[0129]
[0130] in
[0131] Where: H is the measurement matrix, R is the measurement noise covariance matrix; k(·) represents the intensity function of the clutter; the updated Gaussian component mean Covariance Calculated by the following formulas:
[0132]
[0133]
[0134]
[0135] Track start time
[0136] Rearrange D γ,U,k (x) is the index of the Gaussian component in D γ,U,k (x) Arranged as follows:
[0137]
[0138] in is the measurement number;
[0139] D γ,E,k (x) Rearrange as follows:
[0140]
[0141] in
[0142] Finally, all Gaussian components constitute the updated track probability hypothesis density D at time k γ,k (X):
[0143]
[0144] in
[0145] In (S1.3), the sensor track probability hypothesis density D is updated γ,k The Gaussian component of (x) is pruned to update the track probability hypothesis density D γ,k The low-weight Gaussian component in (x) obtains the sensor's posterior track probability hypothesis density And the corresponding posterior track estimate set, denoted as
[0146] In one embodiment, (S1.3) obtaining the posterior track probability hypothesis density of the sensor The following steps are involved:
[0147] (S1.3.1) Set the Gaussian component removal threshold w P The threshold can generally be set to no more than 10 -4 Here we take w P =10 -5 ,if Weight The corresponding Gaussian component is deleted;
[0148] (S1.3.2) Set the total number threshold of Gaussian components I max , here we take I max =150, if the number of Gaussian components is greater than 150, the Gaussian components with weights after 150 will be eliminated;
[0149] (S1.3.3) Target number estimation Calculated by the following formula:
[0150]
[0151] Sort the weights of the track Gaussian components from large to small, and take the first Gaussian components constitute the estimated track probability hypothesis density as follows:
[0152]
[0153] Assume that the posterior track probability hypothesis density corresponding to the two sensors γ∈{γ1,γ2} is Get the posterior track estimation set at time k, denoted as
[0154] The posterior track probability hypothesis density corresponding to the two sensors γ∈{γ1,γ2} It is expressed by the following formula
[0155]
[0156]
[0157]
[0158] Where γ∈{γ1,γ2}, γ1 and γ2 represent sensor numbers, It represents the mean value of the Gaussian component k of the track probability hypothesis density estimated by sensor γ, where That is, the position of the i-th posterior track in the posterior track estimation set of sensor γ1 at time k in the plane rectangular coordinate system and its corresponding speed The Y-axis position of the i-th posterior track in the plane rectangular coordinate system and its corresponding speed
[0159] Sensors γ1 and γ2 estimate the track probability hypothesis density and The track set corresponding to the Gaussian component is recorded as
[0160] In (S2.5), based on the association cost matrix between sensors γ1 and γ2, the posterior tracks in the posterior track estimation set of sensors γ1 and γ2 are divided into associated tracks and non-associated tracks. Specifically, the method is as follows:
[0161] For the i-th posterior track in the posterior track estimation set of sensor γ1, If a i,j =1, and It means that the i-th posterior track in the posterior track estimation set of sensor γ1 is associated with the j-th posterior track in the posterior track estimation set of sensor γ2, and the i-th posterior track in the posterior track estimation set of sensor γ1 and the j-th posterior track in the posterior track estimation set of sensor γ2 are associated tracks, and the associated track index is obtained at the same time
[0162] For the i-th posterior track in the posterior track estimation set of sensor γ1, If a i,j =1, and This means that the i-th posterior track in the posterior track estimation set of sensor γ1 is a non-correlated track, and the non-correlated track index of sensor γ1 is obtained at the same time
[0163] For the i-th posterior track in the posterior track estimation set of sensor γ1, If a i,j =1, and This means that the jth posterior track in the posterior track estimation set of sensor γ2 is a non-associated track, and the non-track index of sensor γ2 is obtained at the same time.
[0164] in, represents the associated track index, L represents the number of associated tracks; at the same time, represents a set of non-associated tracks, Indicates a non-associative track index.
[0165] According to the track association results, the Gaussian component index numbers can be reorganized. It is expressed as:
[0166]
[0167]
[0168]
[0169] In S3, the probability hypothesis density of the associated track between sensors γ1 and γ2 is fused to obtain the fused associated track probability hypothesis density between sensors γ1 and γ2, including:
[0170] The probability hypothesis density of the associated tracks of different sensors γ∈{γ1,γ2} is expressed as:
[0171]
[0172]
[0173] The fusion-related track probability hypothesis density between the fusion sensors γ1 and γ2 is expressed as:
[0174]
[0175] in:
[0176]
[0177]
[0178] in, and Represents the fusion weight, here we take
[0179] In S4, the erroneous tracks in the unrelated tracks between sensors γ1 and γ2 are eliminated by:
[0180] Eliminate the erroneous tracks whose length is less than the threshold σ, where σ=3. The specific process is as follows:
[0181]
[0182] Culling middle Gaussian component, after removing the Gaussian component, the probability hypothesis density of non-correlated tracks The index number of the Gaussian component in the result is:
[0183]
[0184] in Indicates the number of tracks to be removed.
[0185] In S4, the probability hypothesis density of the non-correlated track after removing the erroneous track is fused with the probability hypothesis density of the fused correlated track to obtain the first fused track probability hypothesis density between sensors γ1 and γ2.
[0186] tidy and The index numbers of the Gaussian components are recorded as:
[0187]
[0188] tidy The index numbers of all Gaussian components in Arrange as
[0189]
[0190] in
[0191] The first fused track probability hypothesis density between sensors γ1 and γ2 and the posterior track probability hypothesis density of the third sensor are fused using the same method as in steps (S2) to (S4) to obtain a second fused track probability hypothesis density This process is repeated until all the posterior track probability hypothesis densities of all sensors are fused to obtain the final fused track probability hypothesis density. for:
[0192]
[0193] Based on the final fused track probability hypothesis density Extract the multi-sensor estimated track state
[0194] Steps (S1) to (S5) are repeated until the tracking is completed.
[0195] In the above embodiment, the posterior track probability hypothesis density of the first sensor and the second sensor is firstly fused to obtain the first fused track probability hypothesis density The first fused track probability hypothesis density and the posterior track probability hypothesis density of the third sensor are fused to obtain the second fused track probability hypothesis density This process is repeated until all the posterior track probability hypothesis densities of all sensors are fused to obtain the final fused track probability hypothesis density. In other embodiments, the posterior track probability hypothesis densities of the Y sensors may be fused in pairs (e.g., the posterior track probability hypothesis densities of the first sensor and the second sensor are fused, the posterior track probability hypothesis densities of the third sensor and the fourth sensor are fused, and so on), and then the fused track probability hypothesis densities obtained by the pairwise fusion are fused in pairs, and so on, until the posterior track probability hypothesis densities of all sensors are fused to obtain the final fused track probability hypothesis density. Of course, those skilled in the art can also obviously obtain other regular fusion methods based on the concept of the present invention, among which the fusion method between the probability hypothesis densities of each track is:
[0196] Constructing an association cost matrix, and dividing the posterior tracks in the posterior track estimation set corresponding to the probability hypothesis density of the two tracks to be fused into associated tracks and non-associated tracks based on the association cost matrix;
[0197] The probability hypothesis density of the associated tracks is fused to obtain the fused probability hypothesis density of the associated tracks;
[0198] The erroneous tracks in the non-associated tracks are eliminated, and the probability hypothesis density of the non-associated tracks after eliminating the erroneous tracks is fused with the probability hypothesis density of the fused associated tracks to obtain the current probability hypothesis density of the fused tracks.
[0199] The implementation methods of the above steps can be implemented by the same methods as in the above embodiments, which will not be described again here.
[0200] In order to better illustrate the technical effect of the present invention, the present invention is further described below in conjunction with simulation experiments:
[0201] The simulation experiment environment is an Intel i7 2.3Hz 8-core CPU processor, and the program is written in Matlab language.
[0202] The method of the present invention is denoted as TPHD-TA-GCI-TPHD, the existing GCI fusion method based on probability hypothesis density (PHD) is denoted as PHD-GCI, the existing weighted arithmetic average (WAA) fusion method based on PHD is denoted as PHD-WAA, and the existing flooding arithmetic average fusion method based on Bernoulli and Bernoulli is denoted as B2B-AA-Flodding. These four fusion methods are compared in simulation experiments.
[0203] Simulation condition settings:
[0204] A square two-dimensional area of size [0,1000]×[0,1000](m) is selected as the monitoring area, and two sensors are used for simultaneous monitoring;
[0205] A total of 10 targets appear during the simulation, and the target tracks are as follows: Figure 3 As shown in , the starting position of the track is marked with a circle, and the starting time and length of the track are marked at the end position of the track;
[0206] The state of the target is represented by in is the horizontal and vertical position of the target, are the lateral and longitudinal velocities of the target;
[0207] Sampling interval τ = 1s, the state transfer matrix F is:
[0208]
[0209] The process noise covariance matrix, Q, is given by:
[0210]
[0211] The initial probability density is:
[0212]
[0213] Where γ∈{γ1,γ2}, the initial weight Start time (set to 1 at the initial moment), initial mean Initial covariance The new target probability hypothesis density D at each subsequent moment γ,B,k|k-1 (x) and D γ,0 (x) the same;
[0214] The measurement matrices of both sensors are:
[0215]
[0216] The measurement noise covariance matrix R = diag([3,3] T );
[0217] The clutter intensity is k k (z) = 1 × 10 -5 (m 2 ) -1 , clutter is evenly distributed in the monitoring area;
[0218] The two sensors have the same detection probability, i.e. And the detection probability is 0.9;
[0219] The target survival probability is
[0220] Simulation results analysis:
[0221] Figure 4 It is a schematic diagram of the track estimation of sensor 1 in the simulation example; Figure 5 It is a schematic diagram of the track estimation of sensor 2 in the simulation example; Figure 6 This is a diagram of the fusion track estimation effect obtained by using the method of the present invention in a simulation example. In the figure, "+" represents sensor measurement. The dotted line represents the actual track of 10 targets, the solid line represents the track of the surviving target at time k, "*" represents the single sensor estimated track, and "o" represents the TPHD-TA-GCI fusion track. The starting position of the track is marked with different numbers. Figure 6 It can be seen that the TPHD filter can output the track in real time; TPHD-TA-GCI fusion can not only improve the track estimation accuracy (closer to the real track and smoother than a single sensor), but also reduce missed detections (for example, it can estimate the track (2) missed by sensor 1 and the track (1,3) missed by sensor 2).
[0222] Figure 7 This is a comparison chart of the GOSPA metric error obtained by integrating the four methods of TPHD-TA-GCI, PHD-GCI, PHD-WAA, and B2B-AA-Flooding in the simulation example. Here, the root mean square (RMS) of the generalized optimal sub-pattern assignment (GOSPA) metric is used for evaluation, and 60 Monte Carlo experiments are averaged. Figure 6 It can be seen that the GOSPA error of TPHD-TA-GCI fusion is significantly smaller than that of PHD-GCI, PHD-WAA and B2B-AA-Flodding, indicating that the tracking performance of TPHD-TA-GCI filter is significantly improved.
[0223] Figure 8 This is a comparison chart of the estimated position errors obtained by fusing the four methods of TPHD-TA-GCI, PHD-GCI, PHD-WAA, and B2B-AA-Flooding in the simulation example. It can be seen from the figure that the position estimation accuracy of TPHD-TA-GCI fusion is the highest.
[0224] Fig. 9 This is a comparison chart of missed detection errors obtained by fusing the four methods of TPHD-TA-GCI, PHD-GCI, PHD-WAA, and B2B-AA-Flooding in the simulation example. It can be seen from the figure that the missed detection target of the TPHD-TA-GCI fusion method is significantly lower than that of other algorithms, indicating that the TPHD-TA-GCI fusion method can give full play to the complementary characteristics of multiple sensors.
[0225] Fig.10 This is a comparison diagram of the wrong target errors obtained by fusing the four methods of TPHD-TA-GCI, PHD-GCI, PHD-WAA, and B2B-AA-Flooding in the simulation example. It can be seen from the figure that the TPHD-TA-GCI fusion method has a smaller wrong target estimation, indicating that the TPHD-TA-GCI fusion method can effectively eliminate erroneous tracks.
[0226] Fig.11 This is a comparison chart of the target number estimation results obtained by fusing the four methods of TPHD-TA-GCI, PHD-GCI, PHD-WAA, and B2B-AA-Flooding in the simulation example; it can be seen from the figure that the TPHD-TA-GCI fusion method has accurate target number estimation performance.
[0227] The running time of PHD-GCI, B2B-AA-Flooding, PHD-WAA, and TA-GCI-TPHD are 0.57, 0.46, 0.23, and 0.10, respectively. It can be seen that the TPHD-TA-GCI fusion method takes much less time than other fusions, indicating that the computational efficiency of the present invention is very high and has a good prospect for engineering application.
[0228] In one embodiment, a multi-sensor multi-target tracking device based on a posteriori track estimation is provided, comprising:
[0229] The first module is used to obtain the posterior track probability hypothesis density of each sensor based on the prior track probability hypothesis density of each sensor by using the track probability hypothesis density filtering method, and obtain the posterior track estimation set corresponding to the posterior track probability hypothesis density of different sensors;
[0230] The second module is used to construct an association cost matrix between different sensors, and divide the posterior tracks in the posterior track estimation sets of different sensors into associated tracks and non-associated tracks based on the association cost matrix between different sensors;
[0231] The third module is used to fuse the probability hypothesis density of the associated tracks between different sensors to obtain the fused associated track probability hypothesis density between different sensors;
[0232] The fourth module is used to eliminate the erroneous tracks in the non-correlated tracks between different sensors, and fuse the probability hypothesis density of the non-correlated tracks after eliminating the erroneous tracks with the probability hypothesis density of the fused correlated tracks to obtain the probability hypothesis density of the fused tracks between different sensors;
[0233] The fifth module is used to extract the multi-sensor estimated track state according to the fusion track probability hypothesis density between different sensors.
[0234] The methods for implementing the functions of the above modules can be implemented using the same methods as in the above embodiments, which will not be described in detail here.
[0235] In this embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as shown in FIG. Fig.12 As shown. The computer device includes a processor, a memory, a network interface and a database connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store sample data. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, the steps of the multi-sensor multi-target tracking method based on a posteriori track estimation in the above-mentioned embodiment are implemented.
[0236] Those skilled in the art will understand that Fig.12 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.
[0237] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the multi-sensor multi-target tracking method based on a posteriori track estimation in the above embodiment when executing the computer program.
[0238] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps of the multi-sensor multi-target tracking method based on a posteriori track estimation in the above embodiment are implemented.
[0239] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).
[0240] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0241] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.
Claims
1. A multi-sensor multi-target tracking method based on a posteriori track estimation, characterized in that: include: Based on the prior track probability hypothesis density of each sensor, the posterior track probability hypothesis density of each sensor is obtained by using the track probability hypothesis density filtering method, and the posterior track estimation set corresponding to the posterior track probability hypothesis density of each sensor is obtained; Constructing a correlation cost matrix between different sensors, and dividing the posterior tracks in the posterior track estimation sets of different sensors into correlated tracks and non-correlated tracks based on the correlation cost matrix between different sensors; The association cost matrix between different sensors is constructed, including: According to the probability hypothesis density of the posterior track of different sensors, the similarity matrix of the posterior track estimation set of different sensors is calculated, where for sensor and , k Time sensor and The similarity matrix of the posterior track estimation set is as follows: Similarity Matrix Middle Row, No. Elements of a column Indicates sensor The first A posteriori tracks and sensors The first The similarity of the posterior tracks, , , and Respectively represent sensors and The number of Gaussian components of the posterior track probability hypothesis density; Calculated by the following formula: in represents the multi-factor track similarity, , Indicates the multi-factor weights set, represents the fuzzy factor set, And calculated by the following formula: in, , , represent k Time sensor The Gaussian component mean of the posterior track probability hypothesis density is k Time sensor The first The position of the X-axis of the posterior track in the plane rectangular coordinate system and its corresponding speed , the position of the Y axis and its corresponding speed ; Representative sensor The Gaussian component mean of the posterior track probability hypothesis density is k Time sensor The first j The position of the X-axis of the posterior track in the plane rectangular coordinate system and its corresponding speed , the position of the Y axis and its corresponding speed ; Construct the association matrix of the posterior track estimation set corresponding to different sensors, set the cost of the association matrix based on the similarity matrix, and use the optimization algorithm to find the association matrix with the minimum cost as the final association cost matrix. and The corresponding association matrix of the posterior track estimate set is as follows: In the correlation matrix Row, No. The elements of the column are , is equal to 0 or 1, , and there is at most 1 in each row and column is equal to 1; Setting up the incidence matrix The corresponding cost is , seeking the association matrix with the minimum cost as the final sensor and The association cost matrix between them; The probability hypothesis density of the associated tracks between different sensors is fused to obtain the fused probability hypothesis density of the associated tracks between different sensors; Eliminate the erroneous tracks among the non-correlated tracks between different sensors, fuse the probability hypothesis density of the non-correlated tracks after eliminating the erroneous tracks with the probability hypothesis density of the fused correlated tracks, and obtain the probability hypothesis density of the fused tracks between different sensors; According to the fusion track probability hypothesis density among different sensors, the multi-sensor estimated track state is extracted.
2. The multi-sensor multi-target tracking method based on a posteriori track estimation according to claim 1, characterized in that: Setting multi-factor weights .
3. The multi-sensor multi-target tracking method based on a posteriori track estimation according to claim 1, characterized in that: Sensor-based and The association cost matrix between sensors and The posterior tracks in the posterior track estimation set are divided into associated tracks and non-associated tracks as follows: For sensors The first posterior tracks, where ,if ,and , then the sensor The first A posteriori tracks and sensors The first A posteriori track association, sensor The first A posteriori tracks and sensors The first The posterior tracks are associated tracks; For sensors The first posterior tracks, where ,if ,and , then the sensor The first The posterior tracks are non-correlated tracks; For sensors The first posterior tracks, where ,if ,and , then the sensor The first The posterior tracks are non-correlated tracks.
4. The multi-sensor multi-target tracking method based on a posteriori track estimation according to claim 1, 2 or 3, characterized in that: The method of eliminating erroneous tracks is to set a threshold and eliminate erroneous tracks whose length is less than the set threshold.
5. A multi-sensor multi-target tracking device based on a posteriori track estimation, characterized in that: include: The first module is used to obtain the posterior track probability hypothesis density of each sensor based on the prior track probability hypothesis density of each sensor by using the track probability hypothesis density filtering method, and obtain the posterior track estimation set corresponding to the posterior track probability hypothesis density of different sensors; The second module is used to construct an association cost matrix between different sensors, and divide the posterior tracks in the posterior track estimation sets of different sensors into associated tracks and non-associated tracks based on the association cost matrix between different sensors; The association cost matrix between different sensors is constructed, including: According to the probability hypothesis density of the posterior track of different sensors, the similarity matrix of the posterior track estimation set of different sensors is calculated, where for sensor and , k Time sensor and The similarity matrix of the posterior track estimation set is as follows: Similarity Matrix Middle Row, No. Elements of a column Indicates sensor The first A posteriori tracks and sensors The first The similarity of the posterior tracks, , , and Respectively represent sensors and The number of Gaussian components of the posterior track probability hypothesis density; Calculated by the following formula: in represents the multi-factor track similarity, , Indicates the multi-factor weights set, represents the fuzzy factor set, And calculated by the following formula: in, , , represent k Time sensor The Gaussian component mean of the posterior track probability hypothesis density is k Time sensor The first The position of the X-axis of the posterior track in the plane rectangular coordinate system and its corresponding speed , the position of the Y axis and its corresponding speed ; Representative sensor The Gaussian component mean of the posterior track probability hypothesis density is k Time sensor The first j The position of the X-axis of the posterior track in the plane rectangular coordinate system and its corresponding speed , the position of the Y axis and its corresponding speed ; Construct the association matrix of the posterior track estimation set corresponding to different sensors, set the cost of the association matrix based on the similarity matrix, and use the optimization algorithm to find the association matrix with the minimum cost as the final association cost matrix. and The corresponding association matrix of the posterior track estimate set is as follows: In the correlation matrix Row, No. The elements of the column are , is equal to 0 or 1, , and there is at most 1 in each row and column is equal to 1; Setting up the incidence matrix The corresponding cost is , seeking the association matrix with the minimum cost as the final sensor and The association cost matrix between them; The third module is used to fuse the probability hypothesis density of the associated tracks between different sensors to obtain the fused associated track probability hypothesis density between different sensors; The fourth module is used to eliminate the erroneous tracks in the non-correlated tracks between different sensors, and fuse the probability hypothesis density of the non-correlated tracks after eliminating the erroneous tracks with the probability hypothesis density of the fused correlated tracks to obtain the probability hypothesis density of the fused tracks between different sensors; The fifth module is used to extract the multi-sensor estimated track state according to the fusion track probability hypothesis density between different sensors.
6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the processor executes the computer program, the steps of the multi-sensor multi-target tracking method based on a posteriori track estimation as claimed in claim 1 are implemented.
7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the multi-sensor multi-target tracking method based on a posteriori track estimation as claimed in claim 1 are implemented.
Citation Information
Patent Citations
Multi-sensor multi-target track association and fusion method based on OSPA(2) distance
CN112484728A
Probabilistic Sampling Method For Track Association
US20200090011A1
