Self-adaptive multi-radar robust track association method based on reference topological characteristics

Abstract

The invention discloses an adaptive multi-radar robust track association method based on reference topological characteristics, and belongs to the field of radar-multi-target tracking fusion systems. In a scene comprising m radars and n observation targets, each radar traverses the n observation targets at the current moment to obtain a reference topological feature RET corresponding to each target; then, traversing each radar, randomly selecting one RET element from any target, calculating the maximum value of the distance between every two elements based on the OSPA distance, and taking the maximum value as a reference distance; distributing each RET element in the reference distance by using a Lagrange relaxation algorithm so as to obtain a distribution matrix; thirdly, calculating an association cost matrix by utilizing the distribution matrix and the reference distance; and finally, processing the association cost matrix by using a threshold filtering method, and distributing by using a Lagrange relaxation algorithm to obtain a track association matrix between targets. The method has a good association effect in a multi-radar association scene with a leak detection condition.

Description

Technical Field

[0001] The invention belongs to the field of radar-multi-target tracking fusion system, and in particular relates to an adaptive multi-radar anti-error track association method based on reference topological features. Background Art

[0002] Track correlation is the core of multi-radar data fusion, aiming to determine whether tracks detected by different radars belong to the same target. In a multi-radar, multi-target tracking system, radars upload information to a fusion center for track correlation and fusion. However, radar measurement errors can cause significant deviations between the measured target and the true track, affecting the correlation results. Therefore, track correlation and error estimation are closely related.

[0003] Previous research has mostly treated track correlation and deviation registration separately. For example, the heuristic algorithm proposed by Stone calculates the correlation of radar tracks and estimates relative deviations, but does not consider azimuth deviations. While the global nearest neighbor (GNP) algorithm does account for azimuth deviations, it is difficult to solve due to its mixed-integer nonlinear programming problem. Traditional track correlation methods rely primarily on the absolute position of the target, but systematic deviations can degrade their performance.

[0004] There are two approaches to address this problem: the first is to combine bias estimation with data association, taking bias into account through the expectation maximization (EM) algorithm or the global closest pattern matching (GNPM) method, but this is computationally complex. The other is to select bias-insensitive statistics for bias-resistant association, such as constructing track vectors or algorithms based on reference topological features. While these methods reduce the impact of bias, they may be limited by tracking performance or empirical settings.

[0005] When correlating tracks from three or more radars, the SD allocation problem must be solved. However, this problem is NP-hard. Most bias-resistant track association algorithms are limited to dual-radar scenarios and cannot solve track association between multiple radars. Lagrangian relaxation algorithms can allocate multidimensional matrices, but in track association scenarios where missed detections exist, the algorithm performs a best-effort allocation on all objects, resulting in forced associations between objects from different sources. This requires improvement.

[0006] The algorithm based on state estimation vector decomposition and cancellation can solve the problem of multi-radar track association, but the association accuracy is affected by the accuracy of target state estimation. Summary of the Invention

[0007] In order to solve the problem of multi-radar track correlation in scenarios with radar deviations and improve the robustness of the correlation algorithm, the present invention proposes an adaptive multi-radar robust track correlation method based on reference topology features (RET). The method has good correlation effect in multi-radar correlation scenarios with missed detections, has strong adaptability to radar system deviations and random noise errors, and has a certain improvement in correlation accuracy compared with the multi-radar correlation algorithm based on vector decomposition cancellation.

[0008] The adaptive multi-radar robust track association method based on reference topology features has the following specific steps:

[0009] Step 1: In a scenario with m radars and n observation targets, each radar traverses the n observation targets at the current moment and obtains the reference topological features RET corresponding to each target;

[0010] For all targets observed by the same radar, for a single target i, the relative position information between it and its k adjacent targets is called the reference topological feature RET of target i; k is the number of adjacent targets set artificially; there are n RETs under the same radar; the number of relative position elements included in each RET under all radars is k.

[0011] The RET obtained by radar s observing target i is:

[0012] represents the relative position vector between target i observed by radar s and the adjacent k-th target; R is the topological radius of the neighboring reference track set, represents the track measurement value of target i observed by radar s; Track measurement value The kth element of ; is the track measurement set of neighboring targets within the topological radius R of target i observed by radar s;

[0013] The n RET sets obtained for each target observed by radar s1 are uniformly expressed as Similarly, the n RET sets obtained for each target observed by radar s2 are uniformly expressed as By analogy, the final RET set of all targets under m radars is

[0014] Step 2: Traverse each radar, select a RET element from any target, and calculate the maximum distance between the two elements based on the OSPA distance as the reference distance

[0015]

[0016] Refers to radars p Observation target i p and the adjacent kth p The relative positions of the targets; Refers to radars q Observation target i q and the adjacent kth q The relative position between targets; the benchmark distance is an m-dimensional matrix.

[0017] Step 3: Use the Lagrange relaxation algorithm to allocate each RET element in the benchmark distance to obtain the allocation matrix

[0018]

[0019] k1,k2,...,k m Represents the sequence number of the element in the reference topology structure RET set; h represents the elements in the benchmark distance matrix The distribution matrix between

[0020] Step 4: Calculate the association cost matrix using the allocation matrix and the benchmark distance

[0021]

[0022] in represents the number of targets in the reference topology set, represents the number of matched logarithms in RET, c represents the set threshold, and 1≤r<∞ represents the OSPA metric order. represents the number of unpaired elements in m RETs, represents the total number of paired and unpaired elements in m RETs;

[0023]

[0024] Step 5: Use threshold filtering method to filter the associated cost matrix After processing, the Lagrangian relaxation algorithm is used for distribution to obtain the track correlation matrix between the targets.

[0025] First, the three-dimensional correlation cost matrix C between the sets RET is calculated based on the OSPA distance;

[0026] in

[0027] If all the track pairs in a row or column are within the threshold value c, then all the track pairs in that row fall outside the association threshold, and the corresponding row targets are considered unassociated. At the same time, the association cost matrix C is threshold filtered, and then the Lagrangian relaxation algorithm is used for identification and judgment;

[0028] The formula for the threshold filtering method is as follows:

[0029]

[0030] The advantages of the present invention are:

[0031] 1) The present invention adopts an adaptive multi-radar robust track association method based on reference topological features. For scenarios where multi-radar track association exists and system deviation exists, a multi-radar association metric is constructed, the calculation of the OSPA distance is corrected based on the system deviation, and the association threshold is improved to an adaptive threshold to achieve adaptive track association.

[0032] 2) This adaptive multi-radar robust track association method, based on reference topological features, improves the multidimensional association cost matrix to address missed detection scenarios. A multidimensional matrix allocation algorithm for this scenario is proposed, utilizing a Lagrangian relaxation algorithm to allocate the association cost matrix and determine track associations. Furthermore, this method does not rely on target tracking or state estimation and is equally applicable in complex motion scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 This is a flow chart of the adaptive multi-radar robust track association method based on reference topology features of the present invention;

[0034] Figure 2 This is a diagram of the radar measurement composition of the present invention;

[0035] Figure 3 For the present invention t =4 reference RET definition example diagram;

[0036] Figure 4 Schematic diagram of the Lagrangian relaxation algorithm framework of the present invention. DETAILED DESCRIPTION

[0037] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0038] The present invention proposes an adaptive multi-radar robust track association method based on reference topological features, such as Figure 1 As shown, the input is the measured coordinates in the local polar coordinate system of the radar at each moment, and the targets observed by all radars are traversed to obtain the RET corresponding to each target. The improved Mahalanobis distance between the RET elements is calculated to obtain the multidimensional matrix The optimal matching relationship between RETs is solved, the matrix is ​​preliminarily allocated to obtain the distribution relationship, and then the OSPA distance and association cost matrix between all target corresponding RETs are calculated. The association cost matrix is ​​multi-dimensionally allocated to finally obtain the track association relationship.

[0039] The specific steps are as follows:

[0040] Step 1: In a scenario with m radars and n observation targets, each radar traverses the n observation targets at the current moment and obtains the reference topological features RET corresponding to each target;

[0041] For all targets observed by the same radar, for a single target i, the relative position information between it and its k adjacent targets is called the reference topological feature RET of target i; k is the number of adjacent targets set artificially; there are n RETs under the same radar; the number of relative position elements included in each RET under all radars is k.

[0042] The RET obtained by radar s observing target i is:

[0043] represents the relative position vector between target i observed by radar s and the adjacent k-th target; R is the topological radius of the neighboring reference track set, represents the track measurement value of target i observed by radar s; Track measurement value The kth element of ; is the track measurement set of neighboring targets within the topological radius R of target i observed by radar s;

[0044] The n RET sets obtained for each target observed by radar s1 are uniformly expressed as Similarly, the n RET sets obtained for each target observed by radar s2 are uniformly expressed as By analogy, the final RET set of all targets under m radars is

[0045] Step 2: Traverse each radar, select a RET element from any target, and calculate the maximum distance between the two elements based on the OSPA distance as the reference distance

[0046]

[0047] Refers to radars p Observation target i p and the adjacent kth p The relative positions of the targets; Refers to radars q Observation target i q and the adjacent kth q The relative position between targets; the benchmark distance is an m-dimensional matrix.

[0048] Step 3: Use the Lagrange relaxation algorithm to allocate each RET element in the benchmark distance to obtain the allocation matrix

[0049]

[0050] k1,k2,...,k m Represents the sequence number of the element in the reference topology structure RET set; h represents the elements in the benchmark distance matrix The distribution matrix between

[0051] Step 4: Calculate the association cost matrix using the allocation matrix and the benchmark distance

[0052]

[0053] in represents the number of targets in the reference topology set, represents the number of matched logarithms in RET, c represents the set threshold, and 1≤r<∞ represents the OSPA metric order. represents the number of unpaired elements in m RETs, represents the total number of paired and unpaired elements in m RETs;

[0054]

[0055] Step 5: Use threshold filtering method to filter the associated cost matrix After processing, the Lagrangian relaxation algorithm is used for distribution to obtain the track correlation matrix between the targets.

[0056] First, the three-dimensional correlation cost matrix C between the sets RET is calculated based on the OSPA distance;

[0057] in

[0058] If all the track pairs in a row or column are within the threshold value c, then all the track pairs in that row fall outside the association threshold, and the corresponding row targets are considered unassociated. At the same time, the association cost matrix C is threshold filtered, and then the Lagrangian relaxation algorithm is used for identification and judgment;

[0059] The formula for the threshold filtering method is as follows:

[0060]

[0061] Example:

[0062] Step 1: Target measurement modeling and RET construction

[0063] Radar measures multiple targets within its detection area. The resulting measurement results are a combination of the target's true position and errors. These errors primarily come from the radar's own system deviation, deviations in the radar's mounting platform, biased radar position measurements, and inaccurate radar timing, collectively referred to as system deviation.

[0064] Random noise during measurement is also a major source of error. It is typically modeled as a Gaussian distribution with a mean of 0 and is independent of the radar's own systematic bias and the target's position. Radar bias can be roughly considered a fixed bias, manifesting as a fixed value across all measurement dimensions.

[0065] The radar measurement of the target is generally modeled as the sum of the target's true state, radar system deviation, and random error, which can be expressed in the global Cartesian coordinate system as follows:

[0066]

[0067] in, is the true state vector; b s is the radar deviation; is subject to zero mean and the covariance matrix is ​​P i A Gaussian random errors are distributed.

[0068] Radar measurements such as Figure 2 As shown in the figure, in the local polar coordinate system with the radar as the origin, the position of the target relative to the radar is expressed as the measurement results of distance and azimuth. Due to the existence of system deviation and random noise errors, the radar's measurement results of the target will deviate from the true position. The measurement results can be expressed as follows:

[0069]

[0070] Among them, r i s and θ i s Indicates the radar's measured position of the target, and Indicates the true position of the target, Δr s and Δθ s Indicates the range and azimuth deviation of the radar, and represent the range and azimuth random noise errors respectively;

[0071]

[0072] Track correlation is to find the matching tracks belonging to the same target in the two radar measurements at that moment, that is, to find the most likely correlation relationship. For multi-radar track correlation scenarios, each radar measures the target range, azimuth, and pitch angle in the array polar coordinate system, which can be expressed as The position coordinates of the radar in the rectangular coordinate system are S = (x s ,y s ). In order to perform unified track association and fusion, the multi-radar track measurements need to be converted to a global Cartesian coordinate system with the fusion center as the coordinate origin. The coordinate conversion formula is as follows:

[0073]

[0074] The measurement noise covariance matrix P will also rotate with the transformation of the measurement coordinate system. The transformed matrix can be approximately expressed as:

[0075]

[0076] Among them, J represents the measurement The Jacobian matrix of the expansion point is:

[0077]

[0078] Traditional track association algorithms often use the absolute position of targets for association. However, this absolute position can be affected by factors such as radar bias, which reduces association accuracy and makes the algorithm lack sufficient stability and robustness. In multi-target observation scenarios, in addition to the absolute positions of the targets themselves, there are also relative positions between targets. This relative position information between targets is called the target topology. Although radar system bias can interfere with the measurement of the absolute position of targets, it has little impact on the relative position information between adjacent targets.

[0079] For an observed single target, the relative position information between it and several adjacent targets is called the reference topology feature (RET) of the target.

[0080] First, for the track measurement of radar S The set of neighbor reference tracks is defined as:

[0081]

[0082] Where R is the topological radius, is a collection The kth element of Representing a collection The cardinality of the target track The set of track measurements within the range of R as the topological radius and centered at elements, namely the target track and its surrounding neighbors goals.

[0083] Define the track measurement of radar S The reference topological characteristics are:

[0084]

[0085] The increase of topological radius R will increase the number of targets in the neighboring reference track set of track measurement, which will greatly increase the complexity of RET's structural information and association calculation. Introduce cardinality constraint: To control the operation time, only consider n t - Tracks of nearby neighbors.

[0086] like Figure 3 As shown, it shows that in the space containing 6 targets, when n t =4. The RET diagram of target 1 and target 6 shows that the RET of different source targets is very different, which is helpful to achieve accurate track association.

[0087] In the presence of systematic bias, RET can be expressed as:

[0088]

[0089] Therefore, the system deviation will not affect RET. Using RET measured by track to correlate the track can achieve the effect of anti-error and avoid the influence of system deviation.

[0090] Step 2: Calculation of track correlation metrics.

[0091] Based on the choice of RET as the base unit for association, the core of the track association problem is to determine the cost calculation method for determining whether different tracks are associated. This transforms the problem of determining association between multiple targets into the problem of finding a local optimal solution among the association costs between different tracks. Traditional track association algorithms typically use the Mahalanobis distance between two measurement points to evaluate the proximity of the measurement points. The smaller the Mahalanobis distance, the greater the probability that the two measurement points belong to the same target.

[0092] Traditional methods for calculating association costs cannot be directly applied to calculating association costs between RETs. Instead, a metric that can measure the similarity between sets is needed. The Optimal Sub-Pattern Matching (OSPA) distance, derived from the Wasserstein distance, seeks to minimize the cost of transitions between two distributions, i.e., the distance between the two distributions.

[0093] In this invention, the OSPA distance is used to measure the distance between RETs. Considering the scenario where m radars observe multiple targets, the The OSPA distance between them is:

[0094]

[0095] Among them, s1,s2,...,s m Indicates the radar serial number, k1, k2, ..., k m Indicates the sequence number of the element in the reference topology set, represents the number of targets in the reference topology set, and 1≤r<∞ represents the OSPA metric order. for The reference distance between them is expressed as Mahalanobis distance. When the distance is greater than the set threshold c,

[0096]

[0097] After calculating the Mahalanobis distance, we need to solve The distribution relationship between them, h represents The distribution matrix between represents the number of matched pairs in RET, represents the number of unpaired elements in m RETs, represents the total number of paired and unpaired elements in m RETs.

[0098]

[0099] Step 3: Improvement of association metrics in the presence of bias.

[0100] In the original OSPA algorithm, the distance between measurement points in a set is calculated using the Mahalanobis distance, and the correlation threshold is set as a parameter adjusted empirically. However, because the correlation threshold is affected by factors such as target distribution and measurement noise, adjusting the threshold solely based on experience is not sufficient to adapt to complex and changing track correlation scenarios. Therefore, it is necessary to improve the OSPA distance calculation method based on scenarios where deviations exist.

[0101] Calculating Mahalanobis distance in original OSPA distance The expression is:

[0102]

[0103] in, Indicates the measurement of radar s1 In the topological structure centered on and track The vector between Similarly. ik,jl represents the estimated error covariance matrix between two pairs of measurements, when the sensor system bias is calibrated, P ik,jl It only contains the measurement noise covariance matrix after target tracking; in the presence of system deviation, it also needs to include the covariance matrix R of the system deviation, which is an empirical value derived from the prior knowledge of the distribution of system deviation.

[0104] P ik,jl =P ik +P jl +R ik +R jl (15)

[0105] The radar measurement noise follows a Gaussian distribution The Jacobian matrices of different measurement points will also be different after coordinate transformation, so the measurement noise covariance matrix will also be different, expressed as:

[0106]

[0107] The measurement noise follows the distribution The measurement noise follows the distribution Therefore, the measurement noise of its track pair follows the distribution Right now:

[0108]

[0109] Similarly, the covariance matrix of the system deviation can be obtained:

[0110]

[0111] When calculating the system deviation covariance matrix, since the system deviation is an inherent property of the radar, there will be a certain correlation between the measurements of the same radar. Therefore, the covariance matrix needs to be corrected by adding the cross-covariance matrix between the measurements of the same radar:

[0112]

[0113] The expansion points of the Jacobian matrix of the cross-covariance of the two target measurements are determined by the two measurement points, and the expression is:

[0114]

[0115] The improved Mahalanobis distance with weighted system deviation Subject to the degree of freedom n = 2 2Distribution. To achieve association, χ 2 The threshold condition of the distribution, wherein the threshold value can be taken as a value corresponding to a confidence level of 0.95, and thus the associated threshold does not need to be manually set and adjusted.

[0116] This section improves the original Mahalanobis distance and the manually adjusted association threshold in the OSPA distance into a bias-corrected Mahalanobis distance and an adaptive association threshold. This can adapt to more association scenarios and reduce the instability of the association effect caused by unknown parameters.

[0117] Step 4: Allocation of multi-dimensional correlation cost matrix.

[0118] This embodiment uses multi-target track correlation of three radars. The resulting allocation problem based on the three-dimensional correlation cost matrix is ​​a typical NP-hard combinatorial optimization problem. A suboptimal solution can be obtained using the Lagrangian relaxation algorithm.

[0119] In the actual measurement process, missed detections often occur, which disrupt the integrity of the match between the observed data and the true target, introducing the risk of false associations and increasing association errors. Ideally, data association is based on the optimal match between the target and the observation. However, when a target is not detected (i.e., a missed detection occurs), its corresponding dimension in the cost matrix lacks a valid match, forcing the allocation algorithm to allocate the matching opportunity for that target to other targets. Furthermore, because allocation algorithms (such as the Hungarian algorithm or Lagrangian relaxation) typically tend to strive for matching, they may still attempt to establish associations even if the cost of some matches is high, further increasing the probability of false matches. Therefore, missed detections not only directly lead to the failure to correctly match targets, but also further exacerbate association errors by introducing false matches.

[0120] To reduce false associations, the present invention performs threshold filtering on the association cost matrix. The core concept of threshold filtering is that if all elements in a dimension of the association cost matrix exceed a set threshold, it indicates that the cost of the target and all candidate matches is too high, meaning that the target corresponding to that dimension has not found a reasonable match and is therefore not associated. To achieve this goal, the present invention further sets all cost values ​​for that dimension to infinity. When the cost of a match is set to infinity, that match will never be selected, completely eliminating the possibility of association with that target during the optimization process.

[0121] First, the three-dimensional correlation cost matrix C between the sets RET is calculated based on the OSPA distance;

[0122] in

[0123] If all track pairs in a row or column are within the threshold value c, then all track pairs in that row fall outside the association threshold, and the corresponding row targets are considered unassociated. At the same time, the following processing is performed on C to facilitate the identification and judgment of the Lagrangian relaxation algorithm.

[0124]

[0125] Assume that the dimension of the three-dimensional association cost matrix is ​​n1*n2*n3. The Lagrange algorithm decomposes the three-dimensional assignment problem into a series of relaxed two-dimensional assignment problems by relaxing some constraints. Introducing Lagrange multipliers Substituting the cost function into the relaxed problem, the cost is minimized. The resulting solution is a lower bound on the true solution, while the feasible solution constitutes an upper bound. The difference between the two is called the approximate dual gap. The algorithm iteratively corrects between the dual solution and the feasible solution by adjusting the Lagrange multiplier, gradually bringing them into agreement. When the two are equal, the solution is considered optimal. If a dual solution also satisfies feasibility, it can be directly determined to be optimal.

[0126] The mathematical model of the 3-D assignment problem can be described as

[0127]

[0128] Where J is the objective function, represents the i-th dimension of the m-th dimension m The cost of pairing elements,

[0129]

[0130] The steps to solve the three-dimensional assignment problem using the Lagrangian relaxation algorithm are as follows:

[0131] (1) Initialization: f dual =-∞,f primal =∞, the number of iterations iter = 0, and the maximum number of iterations maxiter = 100 can be set.

[0132] (2) Calculate the cost of the dual problem:

[0133]

[0134] (3) Solve the dual sub-problem:

[0135]

[0136] like make This is a generalized 2-D assignment problem that can be solved using the Hungarian algorithm.

[0137] (4) Modified Lagrange multiplier: In order to generate a series of dual vectors u(0) →u (1) →...u (l) →...u (*) , (l is the number of iterations), there is

[0138]

[0139] H (0) =I.

[0140] When α=2, Has the best iteration effect, and where q *(l) is the optimal value of the dual solution in the first l iterations, and the parameters a and b satisfy 0.05≤a≤0.3, 1.1≤b≤1.6, and

[0141] (5) Construct a feasible solution: in satisfy

[0142]

[0143]

[0144] (6) Iteration: Improve the quality of the solution.

[0145] f dual =max(f dual ,f d ) (32)

[0146] f primal =min(f primal ,f p ) (33)

[0147] gap=(f primal -f dual ) / |f primal | (34)

[0148] If the relative duality gap is less than 0.01 or the number of iterations is greater than the maximum number of iterations, the algorithm ends; otherwise, it goes to step (2). This achieves the distribution of the three-dimensional correlation cost matrix. For the track association of four or more radars, it is necessary to distribute the four-dimensional or higher correlation cost matrix, which can be solved by introducing more Lagrange multipliers.

[0149] The flowchart of the Lagrangian relaxation algorithm is as follows Figure 4 shown.

[0150] The multi-radar track association algorithm based on RET proposed in this paper mainly includes the operations of distance solution and cost matrix allocation. The computational complexity of the algorithm is

[0151] T RET =O{n m k(1+(n t ) m )} (35)

[0152] Where n is the number of targets, m is the number of radars, k is the number of iterations of the Lagrangian relaxation algorithm, and n t The number of elements in the set of neighbor reference tracks measured by the radar.

Claims

1. An adaptive multi-radar robust track association method based on reference topology features, characterized in that: The specific steps are as follows: Step 1: In a scenario with m radars and n observation targets, each radar traverses the n observation targets at the current moment and obtains the reference topological features RET corresponding to each target; The RET obtained by radar s observing target i is: represents the relative position vector between target i observed by radar s and the adjacent k-th target; R is the topological radius of the neighboring reference track set, represents the track measurement value of target i observed by radar s; Track measurement value The kth element of ; is the track measurement set of neighboring targets within the topological radius R of target i observed by radar s; The n RET sets obtained for each target observed by radar s1 are uniformly expressed as Similarly, the n RET sets obtained for each target observed by radar s2 are uniformly expressed as By analogy, the final RET set of all targets under m radars is Step 2: Traverse each radar, select a RET element from any target, and calculate the maximum distance between the two elements based on the OSPA distance as the reference distance Step 3: Use the Lagrange relaxation algorithm to allocate each RET element in the benchmark distance to obtain the allocation matrix k1,k2,...,k m Represents the sequence number of the element in the reference topology structure RET set; h represents the elements in the benchmark distance matrix The distribution matrix between Step 4: Calculate the association cost matrix using the allocation matrix and the benchmark distance in represents the number of targets in the reference topology set, represents the number of matched pairs in RET, c represents the set threshold, and 1≤r<∞ represents the OSPA metric order; represents the number of unpaired elements in m RETs, represents the total number of paired and unpaired elements in m RETs; Step 5: Use threshold filtering method to filter the associated cost matrix After processing, the Lagrangian relaxation algorithm is used to distribute the targets and obtain the track correlation matrix between the targets; First, the three-dimensional correlation cost matrix C between the sets RET is calculated based on the OSPA distance; in The formula for the threshold filtering method is as follows: If all the values ​​in a row or column are the threshold value c, then the track pairs in that row fall outside the association threshold, and the corresponding row targets are considered unassociated. At the same time, the association cost matrix C is threshold filtered, and then the Lagrangian relaxation algorithm is used for identification and judgment.

2. The method according to claim 1, wherein In the step 1, for all targets observed by the same radar, for a single target i, the relative position information between it and the k adjacent targets is called the reference topological feature RET of target i; k is the number of adjacent targets set artificially; there are n RETs under the same radar; the number of relative position elements included in each RET under all radars is k.

3. The method according to claim 1, wherein In step 2, the base distance calculation formula is: Refers to radars p Observation target i p and the adjacent kth p The relative positions of the targets; Refers to radars q Observation target i q and the adjacent kth q The relative position between the targets.

Images