Multi-target track fusion method based on feature matching

Abstract

The invention discloses a multi-target track fusion method based on feature matching. The method comprises the following steps: firstly, realizing a multi-target tracking trajectory of each radar by using a feature matching GM-PHD filter, and numbering each trajectory; then radar feature information is used for assisting track association, and different radars can give the same number to the same target through track association; and finally, carrying out distributed fusion on the local estimation result of each radar node by adopting a GCI algorithm to obtain a fused multi-target track. The method has the advantages of being low in calculation complexity and good in real-time performance while the estimation precision is guaranteed, the target tracking and information fusion capacity of a multi-radar networking system can be effectively improved, and the method is suitable for the field of multi-radar sensing and cooperative detection with high requirements for real-time performance and reliability.

Description

Technical Field

[0001] This invention belongs to the field of multi-radar sensing and cooperative detection technology, specifically involving a multi-target track fusion method based on feature matching. Background Technology

[0002] With the continuous upgrading of military equipment and informationized combat systems, key technologies such as detection, stealth, and electronic countermeasures are constantly developing, making the combat environment highly dynamic and complex. Traditional single-sensor systems (such as single radars) are limited in terms of detection range, target resolution, anti-jamming performance, and multi-target tracking, making it difficult to meet the demands of the modern battlefield for accurate detection and stable identification.

[0003] To adapt to this trend, radar technology is gradually evolving towards networking and collaboration, enhancing overall detection capabilities through multi-source information sharing and integrated processing. Compared to single-radar systems, multi-radar collaborative systems can significantly enhance target detection accuracy and environmental adaptability by utilizing characteristics such as spatial diversity and frequency diversity.

[0004] Against this backdrop, this invention proposes a multi-target track fusion method based on feature matching to address the problems of track fusion and target association under multi-radar networking conditions. This method achieves unified tracking and identification of multiple targets by matching and fusing track features output from multiple radars, thereby improving the overall detection performance and information utilization efficiency of the system. Summary of the Invention

[0005] To address the problem of multi-target track fusion in multi-radar networking, this invention provides a feature-matching-based multi-target track fusion method. Its core technology involves firstly using a GM-PHD (Gaussian Mixture-Probability Hypothesis Density) filter based on feature matching to achieve multi-target tracking for each radar in the radar network, thereby obtaining the multi-target tracking trajectory for each radar and assigning a number to each trajectory; then, using radar feature information to assist in trajectory association, through which different radars assign the same number to the same target; finally, using the GCI (Generalized Covariance Intersection) algorithm for distributed fusion to obtain the fused multi-target tracks, thus achieving multi-target track fusion in multi-radar networking.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0007] This invention provides a multi-target track fusion method based on feature matching, which includes the following steps:

[0008] Step 1: Multi-target tracking of a single radar: Using the GM-PHD filter with feature matching, multi-target tracking of each radar in the radar network is realized, the multi-target tracking trajectory of each radar is obtained, and each trajectory is numbered.

[0009] Step 2, Feature-assisted trajectory association: Based on the multi-target tracking trajectories with numbers obtained in Step 1, radar feature information is used to assist in trajectory association. Through trajectory association, each radar assigns the same number to the same target.

[0010] Step 3, GCI Distributed Fusion: Based on the trajectory association in Step 2, the GCI algorithm is used for distributed fusion to obtain the fused multi-target tracks, realizing multi-target track fusion under multi-radar networking.

[0011] As a further technical solution of the present invention, the specific method of step 1 is as follows:

[0012] (1) Prediction: The probability hypothesis density of the multi-target posterior at time k is expressed as: Then the density function D of trajectory number l k|k The Gaussian mixture representation of (x,l) is:

[0013]

[0014] in, D represents k|k The existence probability of (x,l) is D. k|k (x,l) represents the probability hypothesis density of target l at time k, and x represents the state of a single target. Let J represent the label space of targets at time k, where l represents the target number, satisfying an ordered and distinct sequence, denoted as l = (k, i), where k represents the target's birth time, i is used to distinguish targets born at the same time, and J... k|k (l) represents the number of Gaussian terms at time k. Let represent the weight of the i-th Gaussian term containing position and feature information at time k. Indicates expectation as The covariance matrix is Gaussian probability density function, This represents the expected value of target l at time k. This represents the covariance matrix of target l at time k;

[0015] The probability hypothesis density prediction value at time k+1 is expressed as:

[0016]

[0017] in, D represents S,k+1|kThe existence probability of (x,l) is D. S,k+1|k (x,l) represents the probability hypothesis density prediction value of the surviving targets at time k+1. express The corresponding probability of existence, Let represent the probability hypothesis density of the newborn target numbered l at time k+1. This represents the label space of the newly generated target at time k+1. D represents k+1|k The existence probability of (x,l) is D. k+1|k (x,l) represents the probability hypothesis density prediction value of target number l at time k+1. D represents the label space of the target at time k+1; S,k+1|k (x,l) The calculation formulas are expressed as follows:

[0018]

[0019] Where p S Indicates the probability of the target's survival. Let l represent the expected survival rate of the target numbered l at time k+1. Let represent the covariance matrix of the surviving target numbered l at time k+1. The calculation formula is as follows:

[0020]

[0021] Where F k Let k represent the target state transition matrix. F represents k The transpose matrix, Q k Let k represent the covariance matrix of the process noise at time k.

[0022] The predicted probability hypothesis density value D for target number l at time k+1 is... k+1|k (x,l) is represented as:

[0023]

[0024] Among them, J k+1|k (l) represents the number of Gaussian terms predicted at time k+1. This represents the predicted weight value of the i-th Gaussian term containing location and feature information at time k+1. This represents the expected predicted value of target number l at time k+1. This represents the predicted value of the target covariance matrix numbered l at time k+1;

[0025] (2) Update: At time k+1, the updated value of the multi-target probability hypothesis density is: Among them, D k+1|k+1 (x,l) represents the probability hypothesis density of target number l at time k+1. D represents k+1|k+1 (x,l) represents the probability of existence of the corresponding element.

[0026] The calculation formula is as follows:

[0027]

[0028] At time k+1, using the measurement point set Z k D k+1|k+1 (x,l) is:

[0029]

[0030] Where z represents the measurement point trace containing target location information, (1-p D )D k+1|k (x,l) represents the probability hypothesis density of all missed targets, p D Indicates the detection probability. This represents the weight of the i-th Gaussian term containing position and feature information at time k+1. This represents the expected value of the target numbered l at time k+1. Let l represent the target covariance matrix at time k+1.

[0031] As a further technical solution of the present invention, in step 1, the weight of the i-th Gaussian term containing position and feature information at time k+1 is... The calculation method is as follows:

[0032]

[0033] Among them, κ k+1 (z)=λ·c(z) represents the probability hypothesis density of clutter at time k+1, c(z) represents the spatial distribution of each clutter point, λ represents the average number of clutter points per frame, and z eq This represents a measurement point trace that contains both the target's location and feature information. Indicates by z eq The matching factor between the expected value of the i-th Gaussian component and the expected value of the i-th Gaussian component is calculated using the following formula:

[0034]

[0035] in, Let z represent the matching factor between z and the expected value of the i-th Gaussian component. and They represent the measurement z eqThe matching factor between the Doppler frequency values ​​and signal-to-noise ratio contained in the data and their corresponding predicted values;

[0036] and The calculation formula is:

[0037]

[0038] Among them, H k+1 Represents the measurement matrix. H represents k+1 The transpose matrix, R k+1 The covariance matrix represents the measurement noise. Let I represent the transformation matrix, and let I represent the identity matrix.

[0039] As a further technical solution of the present invention, the specific method of step 2 is as follows:

[0040] (1) Calculate the Mahalanobis distance of the trajectories: The Mahalanobis distance between radar a trajectory numbered l and radar b trajectory numbered l′ is:

[0041] in, and Let represent the mean and variance of the probability hypothesis density for the a-th radar target with radar number l. and Let represent the mean and variance of the probability hypothesis density for the b-th radar target with radar number l′. express and The average value, where a, b ∈ N, and N is a positive integer. and Let these be the target marking spaces for the a-th and b-th radars, respectively;

[0042] (2) Calculate the confidence weight of the Doppler frequency: The confidence weight of the Doppler frequency of radar track number l and radar track number l′ is fd. l,l′ =|fd a (l)-fd b (l′)|, where fd a (l) and fd b (l′) represent the Doppler frequency values ​​of the target numbered l by radar a and the target numbered l′ by radar b, respectively;

[0043] (3) Calculate the trust weight of the signal-to-noise ratio: The trust weight of the signal-to-noise ratio of radar a track number l and radar b track number l′ is SNR. l,l′ =|SNR a (l)-SNR b (l′)|, where SNRa (l) and SNR b (l′) represent the signal-to-noise ratios of the target numbered l by radar a and the target numbered l′ by radar b, respectively;

[0044] (4) Calculate the optimal trajectory association: trajectory association factor The optimal trajectory association between radar a and radar b for:

[0045] Wherein, the mapping trajectory φ a,b (l) indicates from arrive A kind of correlation trajectory, that is, for all All can be found Correspondingly, Represents all mapped trajectories φ a,b (l) constitutes a set, Represents the relationship between trajectory l and mapped trajectory φ a,b (l) is the trajectory correlation factor, where ∑ represents summation and argmin represents the value of the variable when the expression reaches its maximum value;

[0046] (5) Calculate the optimal correlation matrix: through the optimal correlation between each pair of radars The optimal correlation matrix Ω is obtained. Taking the first radar as a reference, the optimal correlation matrix is... in This indicates that the first radar's trajectory is correlated with its own. express, This indicates the optimal trajectory association between radar 1 and radar 2. This represents the optimal trajectory association between radar 1 and radar N. Through trajectory association, each radar is assigned the same number for the same target.

[0047] As a further technical solution of the present invention, the specific method of step 3 is as follows:

[0048] Based on step 1, the probability hypothesis density D of radar a at time k+1 for target number l is calculated. k+1|k+1 (x,l) and the corresponding existence probability The probability hypothesis density D of radar b for target numbered l′ at time k+1 k+1|k+1 (x,l′) and the corresponding existence probability Then the target probability hypothesis density after fusion and the corresponding probability of existence They are represented as follows:

[0049] Among them, w aw represents the fusion weight of radar a. b This represents the fusion weight of radar b.

[0050] As a further technical solution of the present invention, when adopting a pairwise fusion strategy in step 3, for w a and w b Normalize so that w a +w b =1; When the number of radars exceeds two, fusion is achieved through iterative calculation.

[0051] Compared with the prior art, the present invention, employing the above technical solution, has the following technical effects:

[0052] (1) By introducing feature-assisted trajectory association technology, different radars can assign the same number to the same target, avoiding the problem of inconsistent target numbers in traditional methods and improving the target recognition accuracy and consistency of multi-radar systems.

[0053] (2) The method of the present invention fully considers the balance between computational complexity and accuracy in the algorithm design. By combining feature-assisted association and lightweight filtering strategy, it can achieve high-precision multi-target state estimation with low computational cost, taking into account both real-time performance and accuracy. It is suitable for application scenarios with high real-time requirements in multi-radar networking systems.

[0054] (3) This invention achieves effective fusion and trajectory consistency processing of multi-radar observation data by comprehensively utilizing multi-radar observation information and target feature information. This method can overcome the fusion instability caused by the accumulation of measurement errors, target cross motion, or data redundancy in traditional trajectory fusion algorithms, thereby obtaining more accurate, continuous, and consistent data.

[0055] The target trajectory estimation results provide reliable data support for subsequent situation assessment and decision-making. Attached Figure Description

[0056] Figure 1 This is a structural diagram of the method of the present invention;

[0057] Figure 2 It shows the radar position and target trajectory.

[0058] Figure 3 This is a comparison chart of the total OSPA error between the method of this invention and that of a single radar;

[0059] Figure 4 This is a comparison chart of the tracking performance and real-time performance of the method of this invention and a single radar.

[0060] Figure 5 This is a comparison chart of the total OSPA error of different fusion algorithms under radar networking with different parameters;

[0061] Figure 6 This is a comparison chart of the tracking performance and real-time performance of different fusion algorithms under radar networking with different parameters;

[0062] Figure 7 This is a comparison chart of the total OSPA error of different fusion algorithms under radar networking with the same parameters;

[0063] Figure 8 This is a comparison chart of the tracking and real-time performance of different fusion algorithms under radar networking with the same parameters. Detailed Implementation

[0064] 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 denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0065] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless defined as herein.

[0066] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings:

[0067] This invention patent employs a feature-matching-based multi-target trajectory fusion method to solve the problem of multi-target trajectory fusion in multi-radar networking.

[0068] The feature-matching-based multi-target track fusion method is as follows: First, the GM-PHD filter with feature matching is used to achieve multi-target tracking for each radar in the radar network, thereby obtaining the multi-target tracking trajectory for each radar and assigning a number to each trajectory; then, radar feature information is used to assist in trajectory association, and through trajectory association, different radars are assigned the same number to the same target; finally, the GCI algorithm is used for distributed fusion to obtain the fused multi-target track, thereby realizing multi-target track fusion under multi-radar networking.

[0069] This invention relates to a multi-target track fusion method based on feature matching, such as... Figure 1 As shown, it includes the following steps:

[0070] Step 1: Multi-target tracking of a single radar: Using a feature-matching GM-PHD filter, multi-target tracking is achieved for each radar in the radar network, obtaining the multi-target tracking trajectory for each radar, and assigning a number to each trajectory. The specific method is as follows:

[0071] (1) Prediction: The probability hypothesis density of the multi-target posterior at time k is expressed as: Then the density function D of trajectory number l k|k The Gaussian mixture representation of (x,l) is:

[0072]

[0073] in, D represents k|k The existence probability of (x,l) is D. k|k (x,l) represents the probability hypothesis density of target l at time k, and x represents the state of a single target. Let J represent the label space of targets at time k, where l represents the target number, satisfying an ordered and distinct sequence, denoted as l = (k, i), where k represents the target's birth time, i is used to distinguish targets born at the same time, and J... k|k (l) represents the number of Gaussian terms at time k. Let represent the weight of the i-th Gaussian term containing position and feature information at time k. Indicates expectation as The covariance matrix is Gaussian probability density function, This represents the expected value of target l at time k. Let represent the covariance matrix of target l at time k.

[0074] The probability hypothesis density prediction value at time k+1 is expressed as:

[0075]

[0076] in, D represents S,k+1|k The existence probability of (x,l) is D. S,k+1|k (x,l) represents the probability hypothesis density prediction value of the surviving targets at time k+1. express The corresponding probability of existence, Let represent the probability hypothesis density of the newborn target numbered l at time k+1. This represents the label space of the newly generated target at time k+1. D represents k+1|k The existence probability of (x,l) is D. k+1|k(x,l) represents the probability hypothesis density prediction value of target number l at time k+1. D represents the label space of the target at time k+1. S,k+1|k (x,l) The calculation formulas are expressed as follows:

[0077]

[0078] Where p S Indicates the probability of the target's survival. Let l represent the expected survival rate of the target numbered l at time k+1. Let represent the covariance matrix of the surviving target numbered l at time k+1. The calculation formula is as follows:

[0079]

[0080] Where F k Let k represent the target state transition matrix. F represents k The transpose matrix, Q k Let represent the covariance matrix of the process noise at time k.

[0081] The predicted probability hypothesis density value D for target number l at time k+1 is... k+1|k (x,l) is represented as:

[0082]

[0083] Among them, J k+1|k (l) represents the number of Gaussian terms predicted at time k+1. This represents the predicted weight value of the i-th Gaussian term containing location and feature information at time k+1. This represents the expected predicted value of target number l at time k+1. This represents the predicted value of the target covariance matrix numbered l at time k+1.

[0084] (2) Update: At time k+1, the updated value of the multi-target probability hypothesis density is: Among them, D k+1|k+1 (x,l) represents the probability hypothesis density of target number l at time k+1. D represents k+1|k+1 (x,l) represents the probability of existence.

[0085] The calculation formula is as follows:

[0086]

[0087] At time k+1, using the measurement point set Z k D k+1|k+1 (x,l) is:

[0088]

[0089] Where z represents the measurement point trace containing target location information, (1-p D )D k+1|k (x,l) represents the probability hypothesis density of all missed targets, p D Indicates the detection probability. This represents the weight of the i-th Gaussian term containing position and feature information at time k+1. This represents the expected value of the target numbered l at time k+1. Let l represent the target covariance matrix at time k+1.

[0090] The calculation method is as follows:

[0091]

[0092] Among them, κ k+1 (z)=λ·c(z) represents the probability hypothesis density of clutter at time k+1, c(z) represents the spatial distribution of each clutter point, λ represents the average number of clutter points per frame, and z eq This represents a measurement point trace that contains both the target's location and feature information. Indicates by z eq The matching factor between the expected value of the i-th Gaussian component and the expected value of the i-th Gaussian component is calculated using the following formula:

[0093]

[0094] in, Let z represent the matching factor between z and the expected value of the i-th Gaussian component. and They represent the measurement z eq The matching factor between the Doppler frequency values ​​and signal-to-noise ratio contained in the data and their corresponding predicted values.

[0095] and The calculation formula is:

[0096]

[0097] Among them, H k+1 Represents the measurement matrix. H represents k+1 The transpose matrix, R k+1 The covariance matrix represents the measurement noise. Let I represent the transformation matrix, and let I represent the identity matrix.

[0098] Step 2, Feature-Assisted Trajectory Association: Based on the multi-target tracking trajectories with assigned numbers obtained in Step 1, radar feature information is used to assist in trajectory association. Through trajectory association, each radar assigns the same number to the same target. The specific method is as follows:

[0099] The optimal correlation matrix is ​​obtained by using the optimal correlation between pairs of radars. Assuming that the trajectory correlation is performed between radar a and radar b, where a, b∈N and N is a positive integer, the specific steps are as follows:

[0100] (1) Calculate the Mahalanobis distance of the trajectories: The Mahalanobis distance between radar a trajectory numbered l and radar b trajectory numbered l′ is:

[0101] in, and Let represent the mean and variance of the probability hypothesis density for the a-th radar target with radar number l. and Let represent the mean and variance of the probability hypothesis density for the b-th radar target with radar number l′. express and The average value, and These are represented as the target marking spaces for the a-th and b-th radars, respectively.

[0102] (2) Calculate the confidence weight of the Doppler frequency: The confidence weight of the Doppler frequency of radar track number l and radar track number l′ is fd. l,l′ =|fd a (l)-fd b (l′)|, where fd a (l) and fd b (l′) represent the Doppler frequency values ​​of the target numbered l by radar a and the target numbered l′ by radar b, respectively.

[0103] (3) Calculate the trust weight of the signal-to-noise ratio: The trust weight of the signal-to-noise ratio of radar a track number l and radar b track number l′ is SNR. l,l′ =|SNR a (l)-SNR b (l′)|, where SNR a (l) and SNR b (l′) represent the signal-to-noise ratios of the target numbered l by radar a and the target numbered l′ by radar b, respectively.

[0104] (4) Calculate the optimal trajectory association: trajectory association factor The optimal trajectory association between radar a and radar b for:

[0105] Wherein, the mapping trajectory φ a,b (l) indicates from arrive A kind of correlation trajectory, that is, for all All can be found Correspondingly, Represents all mapped trajectories φ a,b (l) constitutes a set, Represents the relationship between trajectory l and mapped trajectory φ a,b The trajectory correlation factor of (l) is represented by the symbol ∑, which indicates summation, and the symbol argmin indicates the value of the variable when the expression reaches its maximum value.

[0106] (5) Calculate the optimal correlation matrix: through the optimal correlation between each pair of radars The optimal correlation matrix Ω is obtained. Taking the first radar as a reference, the optimal correlation matrix is... in This indicates that the first radar's trajectory is correlated with its own. express, This indicates the optimal trajectory association between radar 1 and radar 2. This represents the optimal trajectory association between radar 1 and radar N. Through trajectory association, each radar is assigned the same number for the same target.

[0107] Step 3, GCI Distributed Fusion: Based on the trajectory association in Step 2, distributed fusion is performed using the GCI algorithm to obtain the fused multi-target tracks, realizing multi-target track fusion under multi-radar networking. The specific method is as follows:

[0108] For ease of reading the formula, we still consider the fusion of two radars, a and b. Based on step 1, we calculate the probability hypothesis density D of radar a for target numbered l at time k+1. k+1|k+1 (x,l) and the corresponding existence probability The probability hypothesis density D of radar b for target numbered l′ at time k+1 k+1|k+1 (x,l′) and the corresponding existence probability Then the target probability hypothesis density after fusion and the corresponding probability of existence They are represented as follows:

[0109]

[0110] Among them, w a w represents the fusion weight of radar a.b This represents the fusion weight of radar b.

[0111] When using a pairwise fusion strategy, for w a and w b Normalize so that w a +w b =1; When the number of radars exceeds two, fusion is achieved through iterative calculation.

[0112] The algorithm and processing method of this invention have been verified and have achieved satisfactory application results:

[0113] 1. Experimental conditions:

[0114] In a two-dimensional scene with a monitoring area of ​​[-10001000]′[-10001000] (unit: m), there are a total of 9 targets. Three radars are set up in the simulation experiment, and the radar positions and target trajectories are as follows: Figure 2 As shown.

[0115] The parameters were set as follows in the experiment: sampling interval T = 1s, target survival probability p S =0.95. To evaluate the performance of the method of this invention, 100 Monte Carlo experiments were conducted for each tracking scenario, and the results of the method of this invention were compared with those of a single radar and two other fusion algorithms (PHD-AA fusion algorithm and GA-LMB fusion algorithm). The simulation results were evaluated using OSPA (Optimal Sub-pattern Aggignment) performance and runtime. The hardware environment for the simulation experiment was an Intel(R) Core(TM) i9-10900 CPU and 8GB RAM, and the software platform was a 64-bit Windows 10 operating system and Matlab R2021a.

[0116] 2. Simulation content:

[0117] Experiment 1: Compare the performance of the method of this invention with that of a single radar. The detection probability of radar 1 is set to P. D,1 =0.9, clutter rate is set to λ1=20; detection probability of radar 2 is set to P D,2 =0.85, clutter rate is set to λ2 = 20; detection probability of radar 3 is set to P D,3 =0.9, clutter rate is set to λ3=40. All three radars use feature-matched GM-PHD filters, and the method of this invention fuses the tracks of the three radars. Figure 3 This is a comparison of the total OSPA error of the method of this invention with that of a single radar. Figure 4 This is a comparison of the tracking performance and real-time performance of the method of this invention with that of a single radar.

[0118] Experiment 2: Compare the performance of the proposed method with two other fusion algorithms when the parameters of the three radars are set differently. The detection probability of radar 1 is set to P. D,1 =0.9, clutter rate is set to λ1=20; detection probability of radar 2 is set to P D,2 =0.85, clutter rate is set to λ2 = 20; detection probability of radar 3 is set to P D,3 =0.9, clutter rate is set to λ3=40. Figure 5 This is a comparison of OSPA errors of different fusion algorithms under radar networking with different parameters. Figure 6 This is a comparison of the tracking and real-time performance of different fusion algorithms under radar networking with different parameters.

[0119] Experiment 3: With the parameters of the three radars set identically, compare the performance of the method of this invention with two other fusion algorithms. The detection probability is set to P for all three. D,1 =P D,2 =P D,3 =0.85, and the clutter ratios are all set to λ1=λ2=λ3=40. Figure 7 This is a comparison of OSPA errors of different fusion algorithms under radar networking with the same parameters. Figure 8 This is a comparison of the tracking and real-time performance of different fusion algorithms under radar networking with the same parameters.

[0120] 3. Simulation Result Analysis:

[0121] from Figure 3 and Figure 4 It can be seen that the OSPA total error, OSPA position error, and OSPA potential estimation error of the method of the present invention are all lower than those of a single radar. The total OSPA error of the method of the present invention is 53.27% lower than that of radar 1, 61.18% lower than that of radar 2, and 57.36% lower than that of radar 3. In other words, the tracking performance of the method of the present invention is better than that of a single radar. Since the method of the present invention requires the fusion of the tracks of the three radars, the running time is slightly longer than that of a single radar.

[0122] from Figure 5 and Figure 7 It can be seen that the total OSPA error of the method of the present invention is lower than that of other fusion algorithms, especially when the target appears at t=1s, t=20s, t=40s, t=60s, t=80s and when the target disappears at t=70s.

[0123] from Figure 6It can be seen that the total OSPA error of the method of this invention is 35.28% lower than that of the PHD-AA fusion algorithm and 11.73% lower than that of the GA-LMB fusion algorithm. In other words, the tracking performance of the method of this invention is improved by 35.28% compared to the PHD-AA fusion algorithm and by 11.73% compared to the GA-LMB fusion algorithm. The running time of the method of this invention is shorter than that of the PHD-AA and GA-LMB algorithms. The time cost of the method of this invention is reduced by 94.17% compared to the PHD-AA fusion algorithm and by 69.57% compared to the GA-LMB fusion algorithm. In other words, the real-time performance of the method of this invention is improved by 94.17% compared to the PHD-AA fusion algorithm and by 69.57% compared to the GA-LMB fusion algorithm.

[0124] from Figure 8 It can be seen that the total OSPA error of the method of this invention is 40.53% lower than that of the PHD-AA fusion algorithm and 32.41% lower than that of the GA-LMB fusion algorithm. In other words, the tracking performance of the method of this invention is improved by 40.53% compared to the PHD-AA fusion algorithm and by 32.41% compared to the GA-LMB fusion algorithm. The running time of the method of this invention is shorter than that of the PHD-AA and GA-LMB algorithms. The time cost of the method of this invention is reduced by 95.38% compared to the PHD-AA fusion algorithm and by 71.43% compared to the GA-LMB fusion algorithm. In other words, the real-time performance of the method of this invention is improved by 95.38% compared to the PHD-AA fusion algorithm and by 71.43% compared to the GA-LMB fusion algorithm.

[0125] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any transformations or substitutions that can be conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A multi-target track fusion method based on feature matching, characterized in that, The method comprises the following steps: Step 1, multi-target tracking of a single radar: a GM-PHD filter based on feature matching is used to realize multi-target tracking of each radar in the radar network, to obtain multi-target tracking trajectories of each radar, and to number each trajectory; Step 2, feature-assisted trajectory association: based on the numbered multi-target tracking trajectories of each radar obtained in step 1, radar feature information is used to assist trajectory association, and each radar is given the same number for the same target through trajectory association; Step 3, GCI distributed fusion: based on the trajectory association of step 2, GCI algorithm is used for distributed fusion to obtain fused multi-target trajectories, and multi-target trajectory fusion under multi-radar networking is realized.

2. The multi-target track fusion method based on feature matching according to claim 1, characterized in that, The specific method of step 1 is as follows: (1) Prediction: The multi-target posterior probability hypothesis density at time k is denoted as: The density function D k|k The Gaussian mixture representation of (x, l) is wherein, represents D k|k (x, l) corresponding existence probability, D k|k (x, l) represents the probability hypothesis density of the target numbered l at time k, x represents the single-target state, represents the label space of the target at time k, l represents the number of the target, satisfies an ordered sequence and is mutually different, is represented as l=(k,i), wherein k is used to represent the target birth time, and i is used to distinguish the targets born at the same time, J k|k (l) represents the number of Gaussian terms at time k, represents the weight of the i-th Gaussian term containing position and feature information at time k, represents the expectation is the covariance matrix is a Gaussian probability density function, represents the expectation of the target numbered l at time k, represents the covariance matrix of the target numbered l at time k; The probability hypothesis density prediction value at the k+1 moment is represented as: wherein, denotes D S,k+1|k (x, l) the corresponding existence probability, D S,k+1|k (x, l) denotes the probability hypothesis density prediction value of the surviving target group at time k+1, denotes the corresponding existence probability, denotes the probability hypothesis density of the newly born target numbered l at time k+1, denotes the label space of the newly born target at time k+1, denotes D k+1|k (x, l) the corresponding existence probability, D k+1|k (x, l) denotes the probability hypothesis density prediction value of the target numbered l at time k+1, denotes the label space of the target at time k+1; D S,k+1|k (x, l), The calculation formulas of and are respectively represented as: where p s denotes the target survival probability, denotes the expectation of the surviving target numbered / at time k + 1, denotes the covariance matrix of the surviving target numbered / at time k + 1, The calculation formula of p is as follows: where F k denotes the target state transition matrix at time k, denotes the transpose of F k , Q k denotes the covariance matrix of the process noise at time k; a target probability hypothesis density prediction value D numbered l at time k+1 k+1|k (x, l) is expressed as: wherein J k+1|k (l) denotes the number of Gaussian terms predicted at time k+1, denotes the predicted value of the weight of the i-th Gaussian term containing position and feature information at time k+1, denotes the predicted value of the target expectation numbered l at time k+1, denotes the predicted value of the target covariance matrix numbered l at time k+1; (2) Update: At time k+1, the multi-target probability hypothesis density update value is: where D k+1|k+1 (x, l) represents the probability hypothesis density of the target numbered l at time k+1, represents the existence probability of D k+1|k+1 (x, l) respectively. The calculation formula is as follows: At the k+1 time, the measurement point set Z k D is obtained k+1|k+1 (x, l) is: where z represents the measurement point track containing target position information, (1-p D )D k+1|k (x, l) represents the probability hypothesis density composed of all missed targets, p D represents the detection probability, represents the weight of the i-th Gaussian term containing position and feature information at k+1 time, represents the target expectation numbered l at k+1 time, represents the target covariance matrix numbered l at k+1 time.

3. The multi-target track fusion method based on feature matching according to claim 2, characterized in that, weight of the i-th Gaussian term containing position and feature information at time k+1 The calculation method is as follows: where κ k+1 (z) = λ · c(z) represents the probability hypothesis density of clutter at time k + 1, c(z) represents the spatial distribution of each clutter point, λ represents the average number of clutter points per frame, and z eq represents the measurement point track containing the position information and feature information of the target, represents the measurement point track containing the position information and feature information of the target, eq is the matching factor between the expected value of the i-th Gaussian component and the measurement point track, and the calculation formula is as follows: wherein, represents a matching factor between z and the expected value of the i-th Gaussian component, and respectively represent the measured z eq Doppler frequency value and signal-to-noise ratio contained in the z and the corresponding predicted value; and The calculation formula is: where H k+1 denotes the measurement matrix, denotes H k+1 the transpose matrix, R k+1 denotes the covariance matrix of the measurement noise, denotes the conversion matrix, and I denotes the identity matrix.

4. The multi-target track fusion method based on feature matching according to claim 1, characterized in that, The specific method of step 2 is as follows: (1) Calculate the Mahalanobis distance of the trajectory: the Mahalanobis distance of the radar a trajectory number l and the radar b trajectory number l' is: wherein, and denote the probability hypothesis density mean and variance of the lth target for the ath radar, and denote the probability hypothesis density mean and variance of the l'th target for the bth radar, denote and the average value of a, b e N, N is a positive integer, and denote the target labeling space for the ath and bth radar, respectively; (2) The trust weight of Doppler frequency: The trust weight of Doppler frequency of radar a track number l and radar b track number l' is fd l,l′ = |fd a (l)-fd b (l'), wherein fd a (l) and fd b (l') respectively represent the Doppler frequency value of the target numbered l of the a-th radar and the target numbered l' of the b-th radar. (3) Calculate the trust weight of signal-to-noise ratio: the trust weight of signal-to-noise ratio of radar a track number l and radar b track number l' is SNR l，l′ = | SNR a (l) - SNR b (l') |, wherein SNR a (l) and SNR b (l') respectively represent the signal-to-noise ratio of the target numbered l of the a-th radar and the target numbered l' of the b-th radar; (4) Computing optimal track association: track association factor Then the optimal track association between radar a and radar b Is: where φ a，b (l) denotes a mapping trajectory from to , i.e. for all there exists such that denotes the set of all mapping trajectories φ a，b (l) and denotes the trajectory association factor between trajectory l and mapping trajectory φ a，b (l), where the symbol ∑ denotes summation and the symbol argmin denotes the value of the variable for which the expression is maximized. (5) Computing the optimal association matrix: through the optimal association between each pair of radars The optimal association matrix Ω is obtained, and the optimal association matrix Ω is obtained with the first radar as the reference Wherein The trajectory association of the first radar to itself is represented by The trajectory association between the first radar and the second radar is represented by The optimal trajectory association between the first radar and the Nth radar is represented by Through trajectory association, each radar gives the same number to the same target.

5. The method for multi-target track fusion based on feature matching according to claim 2, characterized in that, The specific method of step 3 is as follows: The probability hypothesis density D of radar a at time k+1 for the target numbered / is calculated according to step 1 k+1|k+1 (x, / ) and the corresponding existence probability The probability hypothesis density D of radar b at time k+1 for the target numbered / ' is calculated according to step 1 k+1|k+1 (x, / ) and the corresponding existence probability The fused target probability hypothesis density D is then and the corresponding existence probability is respectively where w a represents the fusion weight of radar a, w b represents the fusion weight of radar b.

6. The method for multi-target track fusion based on feature matching according to claim 5, characterized in that, When using the pairwise fusion strategy, w a and w b are normalized such that w a + w b = 1. When the number of radars exceeds two, the fusion is realized through iterative calculation.

Images