A multi-platform tracking method based on augmented Lagrange and probabilistic relaxation assignment
By employing augmented Lagrange and probabilistic relaxation assignment methods, the slow convergence speed and target loss issues of the S-dimensional assignment problem in multi-platform passive tracking systems are resolved, achieving more efficient target tracking and track maintenance.
Patent Information
- Application Number
- CN202611251836.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-18
- Publication Date
- 2026-11-10
AI Technical Summary
In existing multi-platform passive tracking systems, the solution to the S-dimensional allocation problem has a slow convergence speed and is prone to target loss and track swapping in complex scenarios.
We employ an augmented Lagrange and probabilistic relaxation assignment approach. By updating the Lagrange multipliers through the augmented Lagrange dual ascent mechanism and combining adaptive adjustment of the penalty parameter with Gibbs weight calculation, we decompose the S-dimensional assignment problem into a two-dimensional sub-problem, outputting multi-dimensional association probabilities and Top-K association hypotheses.
It improves convergence speed, enhances numerical robustness, and significantly improves tracking accuracy and robustness in target intersection and low detection probability scenarios, while avoiding target loss and track swapping.
Smart Images

Figure CN123067572A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of multi-platform, multi-target pure orientation tracking technology, specifically involving a multi-platform tracking method based on augmented Lagrange and probabilistic relaxation assignment. Background Technology
[0002] Multi-platform passive tracking technology refers to the use of multiple observation platforms equipped with passive sensors to detect, estimate the state of, and maintain the track of multiple moving targets by receiving radiation or reflection signals from targets without actively emitting signals. Since passive sensors often only provide angular information, a single platform typically cannot independently complete target localization and tracking. Therefore, it is necessary for multiple platforms to work together, improving tracking performance through data correlation and information fusion.
[0003] In multi-platform passive tracking systems, data correlation is one of the core challenges. Since each platform may acquire multiple measurements (including real target measurements and false alarms) at each sampling time, determining which measurements from different platforms originate from the same target directly impacts the accuracy of subsequent target state estimation. When the system contains S platforms, this problem can be modeled as an S-dimensional assignment (SD Assignment) problem, which involves finding the minimum cost assignment combination that satisfies the constraints within an S-dimensional cost tensor. The standard form of the S-dimensional assignment problem is integer linear programming. When the number of platforms is large or the measurements are dense, its solution complexity increases exponentially with the dimension, making it an NP-hard combinatorial optimization problem that is difficult to solve directly and accurately.
[0004] To address the data correlation problem in multi-platform passive tracking, existing technologies can be mainly categorized as follows:
[0005] The first category is greedy association methods based on Nearest Neighbor (NN) and Global Nearest Neighbor (GNN). These methods pair measurements with targets or measurements with each other at each time step according to the nearest distance criterion. They have advantages such as simple computation and good real-time performance, and are practical in scenarios with sparse targets and little clutter. However, the nearest neighbor method only considers locally optimal pairings and lacks the ability to optimize the global association structure. When multiple targets intersect, measurements are dense, or there are multiple association assumptions with similar costs, the nearest neighbor method is prone to making incorrect decisions, leading to target loss, track swapping, or false track initialization. For S-dimensional association problems in multi-platform passive tracking, the nearest neighbor method often degenerates into pairwise associations, making it difficult to handle complex constraints between multiple platforms.
[0006] The second category is probabilistic association-based methods, mainly including Joint Probabilistic Data Association (JPDA) and Multiple Hypothesis Tracking (MHT). JPDA calculates the association probability between measurements and targets and updates the target state through weighted fusion, which can handle association uncertainty to some extent. However, JPDA requires enumerating all possible association events, and its computational cost increases exponentially with the number of measurements and targets, making it difficult to apply in real time when there are many platforms and dense measurements. MHT maintains multiple association hypothesis trees, preserving various possibilities of historical association decisions, and can theoretically achieve better association results. However, the number of hypotheses increases exponentially over time, requiring pruning and merging strategies for practical application. It is prone to losing true association paths due to premature pruning or causing computational costs to explode due to insufficient pruning. In addition, JPDA and MHT are mainly oriented towards a two-dimensional association framework between targets and measurements, lacking effective integration with S-dimensional allocation relaxation theory, making it difficult to directly utilize multi-dimensional cost structures for efficient probabilistic inference.
[0007] The third category is optimization methods based on Multi-Dimensional Assignment (MDA). These methods model data association in multi-platform passive tracking as an S-dimensional assignment problem, seeking the association combination with the lowest global cost through optimization. This is currently the mainstream technical approach for handling multi-platform passive association problems. A common approach to solving the S-dimensional assignment problem is Lagrangian relaxation: retaining the first two dimensions of constraints to maintain the two-dimensional assignment structure, and introducing Lagrange multipliers to relax the constraints from the 3rd to the Sth dimensions, thereby reducing the original problem to a polynomial-time solvable two-dimensional subproblem. The optimal solution of the original problem is then gradually approximated by iteratively updating the Lagrange multipliers. However, existing Lagrange relaxation methods generally use a subgradient ascent method to update the multipliers, and the step size is usually determined based on a fixed decay rule or a dual gap heuristic estimation. This mechanism has the following drawbacks: First, the subgradient method has a convergence speed that is only sublinear, requiring many iterations and resulting in slow convergence. Second, the step size is sensitive; too large a step size can easily lead to oscillations, while too small a step size can easily cause stagnation. Third, it directly outputs a single allocation result, and when the costs of multiple association assumptions are similar, forcibly selecting a single optimal solution will result in the loss of crucial association uncertainty information. These drawbacks make the traditional Lagrange relaxation method prone to target loss and track swapping problems in complex scenarios such as low signal-to-noise ratio, target convergence, and dense clutter. Summary of the Invention
[0008] This invention addresses the problems of slow convergence speed and target loss and track swapping in complex scenarios of existing S-dimensional allocation problem solving methods. It proposes a multi-platform tracking method based on augmented Lagrange and probabilistic relaxation allocation.
[0009] The technical solution adopted by this invention to solve the above-mentioned technical problems is: a multi-platform tracking method based on augmented Lagrange multiplication and probabilistic relaxation assignment, the method specifically including the following steps:
[0010] Step 1: Obtain The measurement data received by the sensor platform at the current sampling time, of which the first sensor platform will receive the measurement data at the current sampling time. The number of measurements received by the sensor platform at the current sampling time is denoted as . , ;
[0011] Then according to The measurement data received by each sensor platform at the current sampling time is used to construct... Dimensional cost tensor, and according to S-dimensional cost tensor construction problem:
[0012] Step 2: Initialize the number of iterations ;
[0013] Step 3: Perform Lagrangian relaxation on the constructed S-dimensional allocation problem, decomposing it into a two-dimensional allocation subproblem, and obtaining the result of the two-dimensional allocation subproblem in the th... The cost corresponding to the next iteration;
[0014] Step 4: Update the Lagrange multiplier vectors using the augmented Lagrange duality ascent mechanism and penalty parameters to obtain the cost of the measurement combination that satisfies all dimensional constraints. ;
[0015] Step 5: Calculate the total constraint violation based on the constraint violation of each measurement, and adaptively adjust the penalty parameters according to the shrinkage characteristics of the total constraint violation.
[0016] Step Six: Based on the two-dimensional allocator problem in the... The cost corresponding to the next iteration and Calculate the normalized dual gap, and determine whether the convergence condition of the augmented Lagrange algorithm is met based on the normalized dual gap and the total constraint violation. The convergence condition of the augmented Lagrange algorithm is to satisfy condition (1) or condition (2):
[0017] Condition (1) Normalized dual gap converges;
[0018] Condition (2) The total constraint violation degree converges;
[0019] If the convergence condition of the augmented Lagrange algorithm is met, then step seven is executed using the final Lagrange multiplier vectors of each sensor platform.
[0020] If the convergence condition of the augmented Lagrange algorithm is not met, then reconstruct the algorithm. Dimensional cost tensor, let Return to step three;
[0021] Step 7: Use the final Lagrange multiplier vectors from each sensor platform to construct the last... The cost tensor is decoupled and the reduced cost tensor is calculated.
[0022] Step 8: Adaptively select temperature parameters based on the statistical distribution of the reduced cost tensor. Then based on temperature parameters Compute set The Gibbs weights of the measurement combinations corresponding to each element in the dataset;
[0023] Step 9: Perform alternating normalization on the Gibbs weights dimension by dimension to obtain the joint probability tensor;
[0024] Step 10: Sum the joint probability tensor obtained in Step 10 along the complementary dimensions of each dimension to obtain the marginal association probability of each dimension, and generate the Top-K association hypothesis with the highest probability based on the marginal association probability of each dimension and the bundle search.
[0025] Step 11: Build a tracker consisting of JPDA and Kalman filter, and obtain the target coordinates based on weighted least squares method and Top-K correlation assumption. Input the target coordinates into the tracker to update the target position.
[0026] Continuous tracking of the target is achieved by updating the target position at each sampling time.
[0027] Furthermore, the S-dimensional allocation problem is as follows:
[0028]
[0029] in, This indicates the number of measurements received by the first sensor platform at the current sampling time. This indicates the number of measurements received by the second sensor platform at the current sampling time. This indicates the number of measurements received by the third sensor platform at the current sampling time. Indicates the first The number of measurements received by each sensor platform at the current sampling time. Indicates the first The number of measurements received by each sensor platform at the current sampling time. This indicates the number of samples received by the first sensor platform at the current sampling time. The first measurement, the second sensor platform received the first measurement at the current sampling time. The first measurement, ..., the first The sensor platform receives the first data at the current sampling time. The measurement combination consisting of several measurements comes from the negative log-likelihood cost of the same objective; Assign elements to the binary matrix;
[0030] .
[0031] Furthermore, the specific process of step two is as follows:
[0032] Using Lagrange multiplier vectors The S-th constraint set of the relaxation S-dimensional assignment problem is obtained. Dimensional allocator problem:
[0033]
[0034] in, For the first The S-dimensional Lagrange multiplier vector after the nth iteration update , For the first The first measurement received by the sensor platform at the current sampling time is in the... The Lagrange multipliers updated in the next iteration For the first The second measurement received by the sensor platform at the current sampling time is in the... The Lagrange multipliers updated in the next iteration For the first The sensor platform receives the first data at the current sampling time. The measurement was taken at the first The Lagrange multipliers updated in the next iteration;
[0035] Similarly, we continue to use the Lagrange multiplier vectors to successively relax the S-dimensional allocation problem. constraint set, the first The constraint set, ..., the third constraint set, decompose the constructed S-dimensional allocation problem into the following two-dimensional allocation subproblems:
[0036]
[0037] in, The cost of the two-dimensional allocator problem;
[0038] Constructing new binary correlation variables :
[0039]
[0040] And define a new cost tensor for:
[0041] Will and Substituting this into the two-dimensional allocator problem, we obtain the simplified two-dimensional allocator problem:
[0042]
[0043] Solving the two-dimensional allocator problem yields relaxed solutions and their corresponding costs. .
[0044] Furthermore, the specific process of step four is as follows:
[0045] Step 4. Initialization ;
[0046] Step 4.2: Apply the relaxation solution to the first step. Constraint set determination The cost of the dimension allocation problem and the dimensionality allocation problem The sensor platform receives the first data at the current sampling time. The measurement was taken at the first Constraint violation degree of the next iteration ;
[0047] Step 43: Based on the degree of constraint violation Update # A Lagrange multiplier vector of dimension;
[0048] Step 44, Order Determine whether it satisfies ;
[0049] If satisfied The cost of obtaining a combination of measurements that satisfies all dimensional constraints is... Then proceed to step five;
[0050] If not satisfied Then return to step four two.
[0051] Furthermore, the specific process of step four-three is as follows:
[0052]
[0053] in, For the first The sensor platform receives the first data at the current sampling time. The measurement was taken at the first The Lagrange multipliers updated in the next iteration For the first The sensor platform receives the first data at the current sampling time. The measurement was taken at the first The Lagrange multipliers updated in the next iteration For the first The penalty parameters are updated in the next iteration.
[0054] Furthermore, the specific process of step five is as follows:
[0055]
[0056] in, Indicates the first Total constraint violation rate in each iteration Indicates the first The sensor platform in the first The constraint violation degree vector of the next iteration , Indicates the first The first measurement received by the sensor platform at the current sampling time is in the... The degree of constraint violation in the next iteration. Indicates the first The second measurement received by the sensor platform at the current sampling time is in the... The degree of constraint violation in the next iteration. Indicates the first The sensor platform receives the first data at the current sampling time. The measurement was taken at the first The degree of constraint violation in the next iteration. It is an L2 norm;
[0057] Then according to Adjust penalty parameters:
[0058]
[0059] in, As a penalty factor, Indicates the first Total constraint violation rate in each iteration Indicates the upper limit of the penalty parameter. The shrinkage threshold, For the first The penalty parameters are updated in the next iteration.
[0060] Furthermore, the specific process of step eight is as follows:
[0061]
[0062] in, Let be the set of all finite and non-zero elements in the reduced cost tensor. For set The standard deviation of all elements in the sample. These are the initial temperature parameters;
[0063] Then temperature parameters for:
[0064]
[0065] in, For the final selected temperature parameters, For set The minimum value among all elements in;
[0066] According to temperature parameters Compute set The Gibbs weights of the measurement combinations corresponding to each element in the dataset;
[0067]
[0068] in, Represents a set The measurement combination corresponding to the elements in Gibbs weight.
[0069] Furthermore, the specific process of step nine is as follows:
[0070] Step 1: Initialize the number of iterations Initialize the joint probability tensor , The Gibbs weight tensor constructed based on the results of step nine;
[0071] Step 2: Initialize the dimension;
[0072] Step 3, for the first Normalize each dimension:
[0073]
[0074] in, In order to be in The joint probability tensor obtained in the nth iteration is the th The first dimension One value, In order to be in The joint probability tensor obtained in the nth iteration is the th The first dimension One value, For the joint probability tensor, the first The total number of values corresponding to each dimension In order to be in The joint probability tensor obtained in the nth iteration is the th The first dimension One value;
[0075] Step 4: Determine if the dimensions are satisfied. ;
[0076] If satisfied Then proceed to step 5;
[0077] If not satisfied, then let Return to step 3;
[0078] Step 5, determine the first The joint probability tensor obtained in the second iteration is compared to the first iteration. If the change in the joint probability tensor obtained in the next iteration is less than the set tolerance, then proceed to step 6; otherwise, let... Return to step 2;
[0079] Step 6: Perform global normalization on the joint probability tensor obtained in the last iteration to obtain the final joint probability tensor.
[0080] Furthermore, the specific process of step ten is as follows:
[0081] Step A: Calculate the marginal association probability vector for each dimension:
[0082]
[0083] in, For the first A marginal correlation probability vector with multiple dimensions. For measurement combination The corresponding value in the joint probability tensor;
[0084] Step B: Initialize beam width ,in, For the first The number of measurements from each sensor;
[0085] Step C: Sort the measurements of the first dimension in descending order of marginal association probability, and take the top results from the sorted list. The first measure is selected as a candidate measure; then, the measures of the second dimension are sorted in descending order of marginal association probability, and the top measures are selected from the sorted results. If each measurement is used as a candidate measurement, then the following combinations are generated: Each partial allocation has a heuristic probability that is the product of the marginal correlation probabilities of the two measures in the partial allocation. The allocation with the highest heuristic probability is retained. Each part is allocated;
[0086] Step D: Initialize the dimension ;
[0087] Step E: Arrange the first [item] in descending order of marginal association probability. The measurements in each dimension are sorted, and the top-ranked measurements are selected from the sort results. One measurement was used as a candidate measurement;
[0088] Using the newly extracted front Individual measurements and currently retained The combination of parts allocation results in Each part is allocated, and then from the obtained In the partial allocation, the part with the highest heuristic probability is retained. Each part is allocated;
[0089] Step F: Determine the dimension Is it equal to ;
[0090] like equal Then it will be the one that is ultimately retained. Each assignment is used as a Top-K association hypothesis;
[0091] like Not equal to Then let Return to step E.
[0092] Furthermore, the method for calculating the normalized dual gap is as follows:
[0093]
[0094] in, As an intermediate variable, For the first The cost of a relaxed solution to the two-dimensional assignment problem in the next iteration;
[0095]
[0096] in, For the first The normalized dual gap calculated in the next iteration This indicates taking the absolute value.
[0097] The beneficial effects of this invention are:
[0098] This invention replaces the traditional subgradient multiplier update with augmented Lagrange dual ascent, and replaces the heuristic variable step size with an augmented Lagrange-style update based on a fixed penalty parameter. The multiplier sequence converges to the optimal dual solution at a near-linear rate, overcoming the oscillation and stagnation problems of the subgradient method and improving the convergence speed. Based on the characteristics of constraint violation norm, the penalty parameter is dynamically adjusted. When the constraint is not sufficiently contracted, the penalty parameter is automatically increased to strengthen constraint satisfaction; otherwise, the current level is maintained to avoid over-penalization, significantly improving the numerical robustness under different problem instances. Finally, the decision of a single assignment is relaxed into a continuous probability distribution, outputting the marginal correlation probability, correlation entropy, and Top-K hypotheses for each dimension, providing the tracker with rich probabilistic uncertainty information. In target intersection and low detection probability scenarios, the accuracy and robustness of cooperative tracking are significantly improved, effectively avoiding target loss and track swapping. Attached Figure Description
[0099] Figure 1 This is a flowchart of a multi-platform tracking method based on augmented Lagrange and probabilistic relaxation assignment according to the present invention;
[0100] Figure 2 A flowchart for the generalized Lagrange method;
[0101] Figure 3 A flowchart of the probability relaxation assignment method;
[0102] Figure 4 The image shows the tracking results of the Augmented Lagrange and Probabilistic Relaxation Allocation (ALM-PRA) algorithm of this invention.
[0103] Figure 5 The image shows the tracking results of the traditional Lagrange relaxation algorithm without probabilistic relaxation.
[0104] Figure 6 A comparison of the root mean square errors of the two algorithms under different signal-to-noise ratios;
[0105] Figure 7 A comparison diagram of the generalized optimal submode assignment of the two algorithms under different signal-to-noise ratios. Detailed Implementation
[0106] Specific implementation method one: Combining Figure 1 and Figure 3 This embodiment describes a multi-platform tracking method based on augmented Lagrange multiplication and probabilistic relaxation assignment. The method specifically includes the following steps:
[0107] Step 1: Obtain The measurement data received by the sensor platform at the current sampling time, of which the first sensor platform will receive the measurement data at the current sampling time. The number of measurements received by the sensor platform at the current sampling time is denoted as . , ; and then according to The measurement data received by each sensor platform at the current sampling time is used to construct... Dimensional cost tensor;
[0108] And according to S-dimensional cost tensor construction problem:
[0109]
[0110] in, This indicates the number of measurements received by the first sensor platform at the current sampling time. This indicates the number of measurements received by the second sensor platform at the current sampling time. This indicates the number of measurements received by the third sensor platform at the current sampling time. Indicates the first The number of measurements received by each sensor platform at the current sampling time. Indicates the first The number of measurements received by each sensor platform at the current sampling time. This indicates the number of samples received by the first sensor platform at the current sampling time. The first measurement, the second sensor platform received the first measurement at the current sampling time. The first measurement, ..., the first The sensor platform receives the first data at the current sampling time. The measurement combination consisting of several measurements comes from the negative log-likelihood cost of the same objective; Assign elements to the binary matrix;
[0111]
[0112] Step 2: Initialize the number of iterations ;
[0113] Step 3: Perform Lagrangian relaxation on the constructed S-dimensional allocation problem, decomposing it into a two-dimensional allocation subproblem, and obtaining the result of the two-dimensional allocation subproblem in the th... The cost corresponding to the next iteration;
[0114] Specifically,
[0115] Using Lagrange multiplier vectors The S-th constraint set of the relaxation S-dimensional assignment problem is obtained. Dimensional allocator problem:
[0116]
[0117] in, For the first The S-dimensional Lagrange multiplier vector after the nth iteration update , For the first The first measurement received by the sensor platform at the current sampling time is in the... The Lagrange multipliers updated in the next iteration For the first The second measurement received by the sensor platform at the current sampling time is in the... The Lagrange multipliers updated in the next iteration For the first The sensor platform receives the first data at the current sampling time. The measurement was taken at the first The Lagrange multipliers updated in the next iteration;
[0118] Similarly, we continue to use the Lagrange multiplier vectors to successively relax the S-dimensional allocation problem. constraint set, the first The constraint set, ..., the third constraint set, decompose the constructed S-dimensional allocation problem into the following two-dimensional allocation subproblems:
[0119]
[0120] in, The cost of the two-dimensional allocator problem;
[0121] Constructing new binary correlation variables :
[0122]
[0123] And define a new cost tensor for:
[0124] Will and Substituting this into the two-dimensional allocator problem, we obtain the simplified two-dimensional allocator problem:
[0125]
[0126] Solving the two-dimensional allocator problem yields relaxed solutions and their corresponding costs. .
[0127] Step 4: Update the Lagrange multiplier vectors using the augmented Lagrange dual ascent mechanism to obtain the cost of the measurement combination that satisfies all dimensional constraints. ;
[0128] Specifically,
[0129] Step 4. Initialization ;
[0130] Step 4.2: Apply the relaxation solution to the first step. Constraint set determination The cost of the dimension allocation problem and the dimensionality allocation problem The sensor platform receives the first data at the current sampling time. The measurement was taken at the first Constraint violation degree of the next iteration Constraint violation degree Characterizing the first The sensor platform receives the first data at the current sampling time. The measurement was taken at the first The number of times each iteration is associated;
[0131] Step 43: Based on the degree of constraint violation Update # Lagrange multiplier vectors of dimension:
[0132]
[0133] in, For the first The sensor platform receives the first data at the current sampling time. The measurement was taken at the first The Lagrange multipliers updated in the next iteration For the first The sensor platform receives the first data at the current sampling time. The measurement was taken at the first The Lagrange multipliers updated in the next iteration For the first The penalty parameter is updated in the next iteration;
[0134] It should be noted that the initial values of the Lagrange multipliers and the penalty parameters are obtained through initialization;
[0135] Step 44, Order Determine whether it satisfies ;
[0136] If satisfied The cost of obtaining a combination of measurements that satisfies all dimensional constraints is... Then proceed to step five;
[0137] If not satisfied Then return to step four two.
[0138] Step 5: Calculate the total constraint violation based on the constraint violation of each measurement, and adaptively adjust the penalty parameters according to the shrinkage characteristics of the total constraint violation.
[0139] Specifically,
[0140]
[0141] in, Indicates the first Total constraint violation rate in each iteration Indicates the first The sensor platform in the first The constraint violation degree vector of the next iteration , Indicates the first The first measurement received by the sensor platform at the current sampling time is in the... The degree of constraint violation in the next iteration. Indicates the first The second measurement received by the sensor platform at the current sampling time is in the... The degree of constraint violation in the next iteration. Indicates the first The sensor platform receives the first data at the current sampling time. The measurement was taken at the first The degree of constraint violation in the next iteration. It is an L2 norm;
[0142] Then according to Adjust penalty parameters:
[0143]
[0144] in, As a penalty factor, , Indicates the first Total constraint violation rate in each iteration Indicates the upper limit of the penalty parameter. The shrinkage threshold, For the first The penalty parameters are updated in the next iteration.
[0145] Step Six: Based on the cost and Calculate the normalized dual gap, and determine whether the convergence condition of the augmented Lagrange algorithm is met based on the normalized dual gap and the total constraint violation.
[0146] If the convergence condition of the augmented Lagrange algorithm is met, then step seven is executed using the final Lagrange multiplier vectors of each sensor platform.
[0147] If the convergence condition of the augmented Lagrange algorithm is not met, then reconstruct the algorithm. The cost tensor (because constraint violation records the number of times each measure is associated; if a measure is not associated at all, it is considered a false alarm, and the cost of the combination of measures including that measure is set to zero in the reconstructed cost tensor) ),make Return to step three;
[0148] It should be noted that the normalized dual gap is calculated as follows:
[0149]
[0150] in, As an intermediate variable, For the first The cost of a relaxed solution to the two-dimensional assignment problem in the next iteration;
[0151]
[0152] in, For the first The normalized dual gap calculated in the next iteration Indicates taking the absolute value;
[0153] The augmented Lagrange algorithm is considered to have met the convergence condition when either condition (1) or condition (2) is satisfied:
[0154] Condition (1) Normalized dual gap converges;
[0155] Condition (2) The total constraint violation degree converges.
[0156] Step 7: Use the final Lagrange multiplier vectors from each sensor platform to construct the last... The cost tensor is decoupled and the reduced cost tensor is calculated.
[0157] Specifically, the reduction cost tensor is calculated using broadcast subtraction.
[0158] Step 8: Adaptively select temperature parameters based on the statistical distribution of the reduced cost tensor. Then based on temperature parameters Compute set The Gibbs weights of the measurement combinations corresponding to each element in the dataset;
[0159] Specifically,
[0160]
[0161] in, To reduce all finite (i.e. non-finite) values in the cost tensor A set consisting of elements whose values are non-zero. For set The standard deviation of all elements in the sample. These are the initial temperature parameters;
[0162] At the same time, a lower limit is set to prevent numerical problems, i.e., to constrain the temperature parameter. for:
[0163]
[0164] in, For the final selected temperature parameters, For set The minimum value among all elements in;
[0165] According to temperature parameters Compute set The Gibbs weights of the measurement combinations corresponding to each element in the dataset;
[0166]
[0167] in, Represents a set The measurement combination corresponding to the elements in Gibbs weight;
[0168] Step 9: Use the Sinkhorn-Knopp iterative algorithm to perform alternating normalization of the Gibbs weights dimension by dimension, and iteratively project them onto the multidimensional allocation constraint set to obtain a joint probability tensor that satisfies the marginal sum of each dimension not exceeding 1.
[0169] Specifically,
[0170] Step 1: Initialize the number of iterations Initialize the joint probability tensor , The Gibbs weight tensor constructed based on the results of step nine;
[0171] Step 2: Initialize the dimension The total number of dimensions is equal to the number of sensors;
[0172] Step 3, for the first Normalize each dimension:
[0173]
[0174] in, In order to be in The joint probability tensor obtained in the nth iteration is the th The first dimension One value, In order to be in The joint probability tensor obtained in the nth iteration is the th The first dimension One value, For the joint probability tensor, the first The total number of values corresponding to each dimension ( With the (The number of non-false alarm measurements received by the sensors corresponding to each dimension at the current moment is equal.) In order to be in The joint probability tensor obtained in the nth iteration is the th The first dimension One value;
[0175] Step 4: Determine if the dimensions are satisfied. ;
[0176] If satisfied Then proceed to step 5;
[0177] If not satisfied, then let Return to step 3;
[0178] Step 5, determine the first The joint probability tensor obtained in the second iteration is compared to the first iteration. Whether the change in the joint probability tensor obtained in the next iteration is less than the set tolerance (the tolerance is set in this invention). If the change is less than the set tolerance, proceed to step 6; otherwise, let... Return to step 2;
[0179] Step 6: Perform global normalization on the joint probability tensor obtained in the last iteration to obtain the final joint probability tensor.
[0180] Step 10: Sum the joint probability tensor obtained in Step 10 along the complementary dimensions of each dimension to obtain the marginal association probability of each dimension, and generate the Top-K association hypothesis with the highest probability based on the marginal association probability of each dimension and the bundle search.
[0181] Specifically,
[0182] Step A: Calculate the marginal association probability vector for each dimension:
[0183]
[0184] in, For the first The marginal association probability vector of the nth dimension (including the nth dimension) (marginal correlation probability corresponding to each measurement in each dimension). For measurement combination The corresponding value in the joint probability tensor;
[0185] Step B: Initialize beam width ,in, For the first The number of measurements from each sensor;
[0186] Step C: Sort the measurements of the first dimension in descending order of marginal association probability, and take the top results from the sorted list. The first measure is selected as a candidate measure; then, the measures of the second dimension are sorted in descending order of marginal association probability, and the top measures are selected from the sorted results. If each measurement is used as a candidate measurement, then the following combinations are generated: Each partial allocation has a heuristic probability that is the product of the marginal correlation probabilities of the two measures in the partial allocation. The allocation with the highest heuristic probability is retained. Each part is allocated;
[0187] Step D: Initialize the dimension ;
[0188] Step E: Arrange the first [item] in descending order of marginal association probability. The measurements in each dimension are sorted, and the top-ranked measurements are selected from the sort results. One measurement was used as a candidate measurement;
[0189] Using the newly extracted front Individual measurements and currently retained The combination of parts allocation results in Each partial allocation (the heuristic probability of the combined partial allocation is the product of the marginal correlation probability of the newly extracted measurement and the heuristic probability of the original partial allocation), and then from the obtained In the partial allocation, the part with the highest heuristic probability is retained. Each part is allocated;
[0190] Step F: Determine the dimension Is it equal to ;
[0191] like equal Then it will be the one that is ultimately retained. Each assignment is used as a Top-K association hypothesis;
[0192] like Not equal to Then let Return to step E.
[0193] Step 11: Establish a tracker consisting of JPDA and Kalman filter, and obtain the target coordinates based on weighted least squares method and Top-K correlation assumption. Input the target coordinates into the tracker to update the target position, and realize continuous tracking of the target based on the target position update results corresponding to each sampling time.
[0194] It should be noted that the covariance matrix of the Kalman filter is based on the correlation entropy. Sure:
[0195]
[0196] in, For the first Victor Marginal correlation probability of a measurement It is the natural logarithm;
[0197] If the correlation entropy exceeds a preset threshold, the covariance matrix of the corresponding fused measurement will be automatically increased to reduce the filter's confidence in ambiguous measurements; if the correlation entropy does not exceed the preset threshold, the covariance matrix does not need to be adjusted.
[0198] Example
[0199] Step 1: Deploy five sensor platforms in a planar coordinate system. Each platform is equipped with a passive sonar for receiving target azimuth measurements. Based on the measurement geometry and signal strength, construct a 5-dimensional cost tensor. Each element in the cost tensor represents a measurement combination. The negative log-likelihood cost from the same objective, where:
[0200] This indicates that the first sensor platform received the first... Azimuth angle measurement This indicates that the second sensor platform received the first... Azimuth angle measurement This indicates that any number of data received by the third sensor platform is... Azimuth angle measurement This indicates that the fourth sensor platform received the first... Azimuth angle measurement This indicates that the fifth sensor platform received the first... Azimuth angle measurement;
[0201] Construct a 5-dimensional allocation problem based on a 5-dimensional cost tensor:
[0202]
[0203] in, This indicates the number of measurements received by the first sensor platform at the current sampling time. This indicates the number of measurements received by the second sensor platform at the current sampling time. This indicates the number of measurements received by the third sensor platform at the current sampling time. This indicates the number of measurements received by the fourth sensor platform at the current sampling time. This indicates the number of measurements received by the 5th sensor platform at the current sampling time. This indicates the number of samples received by the first sensor platform at the current sampling time. The first measurement, the second sensor platform received the first measurement at the current sampling time. The measurement, ..., the 5th sensor platform received the first measurement at the current sampling time. The measurement combination consisting of several measurements comes from the negative log-likelihood cost of the same objective; The elements of a binary matrix are assigned as follows:
[0204]
[0205] That is, when the first sensor platform receives the first sample at the current sampling time... The first measurement, the second sensor platform received the first measurement at the current sampling time. The measurement, ..., the 5th sensor platform received the first measurement at the current sampling time. When a combination of measurements originates from the same target... The value is 1, otherwise, The value is 0;
[0206] Step 2: Initialize the number of iterations ;
[0207] Step 3: Perform Lagrange relaxation on the constructed five-dimensional cost tensor, preserving the constraints of the first two dimensions to ensure the two-dimensional allocation structure. That is, introduce Lagrange multipliers into the constraints of the 3rd, 4th, and 5th dimensions of the 5-dimensional cost tensor, decomposing the constructed five-dimensional allocation problem into a two-dimensional allocator problem with respect to the first two dimensions. The result is the two-dimensional allocator problem at the 3rd, 4th, and 5th dimensions. The cost corresponding to the next iteration;
[0208] Specifically,
[0209] Step 31: Using Lagrange multiplier vectors Relaxing the fifth constraint set of the five-dimensional allocation problem yields a four-dimensional allocation subproblem:
[0210]
[0211] in, For the first The fifth-dimensional Lagrange multiplier vector after the second iteration update , The first measurement received by the fifth sensor platform at the current sampling time is in the... The Lagrange multipliers updated in the next iteration The second measurement received by the fifth sensor platform at the current sampling time is in the... The Lagrange multipliers updated in the next iteration For the fifth sensor platform, the first sensor received at the current sampling time... The measurement was taken at the first The Lagrange multipliers updated in the next iteration;
[0212] Step 3.2: Using Lagrange multiplier vectors Relaxing the fourth set of constraints in the five-dimensional allocation problem yields a three-dimensional allocation subproblem:
[0213]
[0214] in, For the first The fourth-dimensional Lagrange multiplier vector after the next iteration update , The first measurement received by the fourth sensor platform at the current sampling time is in the... The Lagrange multipliers updated in the next iteration The second measurement received by the fourth sensor platform at the current sampling time is in the... The Lagrange multipliers updated in the next iteration For the fourth sensor platform, the first sensor received the data at the current sampling time. The measurement was taken at the first The Lagrange multipliers updated in the next iteration;
[0215] Step 33: Using Lagrange multiplier vectors Relaxing the third set of constraints in the five-dimensional allocation problem yields a two-dimensional allocation subproblem:
[0216]
[0217] in, Let be the third-dimensional Lagrange multiplier vector. , The first measurement received by the third sensor platform at the current sampling time is in the... The Lagrange multipliers updated in the next iteration The second measurement received by the third sensor platform at the current sampling time is in the... The Lagrange multipliers updated in the next iteration For the third sensor platform, the number of samples received at the current sampling time is... The measurement was taken at the first The Lagrange multipliers updated in the next iteration;
[0218] Steps 3 and 4: Construct new binary correlation variables :
[0219]
[0220] Define a new cost tensor for:
[0221] Will and Substituting this into the two-dimensional allocator problem, we obtain the simplified two-dimensional allocator problem:
[0222]
[0223] Solving the two-dimensional allocator problem yields relaxed solutions and their corresponding costs. .
[0224] Step 4, as follows Figure 2 As shown, the cost of updating the Lagrange multiplier vectors using the augmented Lagrange dual ascent mechanism is obtained to satisfy the measurement combination that satisfies all dimensional constraints. ;
[0225] Specifically,
[0226] Step 4. Initialization ;
[0227] Step 4.2: Apply the relaxation solution to the first step. Constraint set determination The cost of the dimension allocation problem and the dimensionality allocation problem The sensor platform receives the first data at the current sampling time. The measurement was taken at the first Constraint violation degree of the next iteration Constraint violation degree Characterizing the first The sensor platform receives the first data at the current sampling time. The measurement was taken at the first The number of times each iteration is associated;
[0228] Step 43: Based on the degree of constraint violation Update # Lagrange multiplier vectors of dimension:
[0229]
[0230] in, For the first The sensor platform receives the first data at the current sampling time. The measurement was taken at the first The Lagrange multipliers updated in the next iteration For the first The sensor platform receives the first data at the current sampling time. The measurement was taken at the first The Lagrange multipliers updated in the next iteration For the first The penalty parameter is updated in the next iteration;
[0231] It should be noted that the initial values of the Lagrange multipliers and the penalty parameters are obtained through initialization;
[0232] Step 44, Order Determine whether it satisfies ;
[0233] If satisfied The cost of obtaining a combination of measurements that satisfies all dimensional constraints is... Then proceed to step five;
[0234] If not satisfied Then return to step four two.
[0235] Step 5: Adaptively adjust the penalty parameters based on the shrinkage characteristics of constraint violation.
[0236] Specifically,
[0237]
[0238] in, Indicates the first Total constraint violation rate in each iteration Indicates the first The sensor platform in the first The constraint violation degree vector of the next iteration , Indicates the first The first measurement received by the sensor platform at the current sampling time is in the... The degree of constraint violation in the next iteration. Indicates the first The second measurement received by the sensor platform at the current sampling time is in the... The degree of constraint violation in the next iteration. Indicates the first The sensor platform receives the first data at the current sampling time. The measurement was taken at the first Constraint violation degree in the next iteration;
[0239] Then according to Adjust penalty parameters:
[0240]
[0241] in, As a penalty factor, , Indicates the first Total constraint violation rate in each iteration Indicates the upper limit of the penalty parameter. The shrinkage threshold, For the first The penalty parameters are updated in the next iteration.
[0242] Step Six: Based on the cost and Calculate the normalized dual gap, and determine whether the convergence condition of the augmented Lagrange algorithm is met based on the normalized dual gap and the total constraint violation.
[0243] If the convergence condition of the augmented Lagrange algorithm is met, then step seven is executed using the final Lagrange multiplier vectors of each sensor platform.
[0244] If the convergence condition of the augmented Lagrange algorithm is not met, then let Return to step three;
[0245] It should be noted that the normalized dual gap is calculated as follows:
[0246]
[0247] in, As an intermediate variable, For the first The cost of a relaxed solution to the two-dimensional assignment problem in the next iteration;
[0248]
[0249] in, For the first The normalized dual gap calculated in the next iteration Indicates taking the absolute value;
[0250] The augmented Lagrange algorithm is considered to have met the convergence condition when either condition (1) or condition (2) is satisfied:
[0251] Condition (1) Normalized dual gap converges;
[0252] Condition (2) The total constraint violation degree converges.
[0253] Step 7: Use the final Lagrange multiplier vectors from each sensor platform to construct the last... The cost tensor is decoupled and the reduced cost tensor is calculated.
[0254] Specifically, the reduction cost tensor is calculated using broadcast subtraction.
[0255] Step 8: Adaptively select temperature parameters based on the statistical distribution of the reduced cost tensor. ;
[0256] Specifically,
[0257]
[0258] in, To reduce all finite (i.e. non-finite) values in the cost tensor A set consisting of elements whose values are non-zero. For set The standard deviation of all elements in the sample. These are the initial temperature parameters;
[0259] Temperature parameters for:
[0260]
[0261] in, For the final selected temperature parameters, For set The minimum value among all elements in;
[0262] According to temperature parameters Compute set The Gibbs weights of the measurement combinations corresponding to each element in the dataset;
[0263] Specifically,
[0264]
[0265] in, Represents a set The measurement combination corresponding to the elements in Gibbs weights:
[0266] Step 9: Use the Sinkhorn-Knopp iterative algorithm to perform alternating normalization of the Gibbs weights dimension by dimension, and iteratively project them onto the multidimensional allocation constraint set to obtain a joint probability tensor that satisfies the marginal sum of each dimension not exceeding 1.
[0267] Specifically,
[0268] Step 1: Initialize the number of iterations Initialize the joint probability tensor ;
[0269] Step 2: Initialize the dimension ;
[0270] Step 3, for the first Normalize each dimension:
[0271]
[0272] in, In order to be in The joint probability tensor obtained in the nth iteration is the th The first dimension One value, In order to be in The joint probability tensor obtained in the nth iteration is the th The first dimension One value, For the joint probability tensor, the first The total number of values corresponding to each dimension In order to be in The joint probability tensor obtained in the nth iteration is the th The first dimension One value;
[0273] Step 4: Determine if the dimensions are satisfied. ;
[0274] If satisfied Then proceed to step 5;
[0275] If not satisfied, then let Return to step 3;
[0276] Step 5, determine the first The joint probability tensor obtained in the second iteration is compared to the first iteration. Whether the change in the joint probability tensor obtained in the next iteration is less than the set tolerance (the tolerance is set in this invention). If the change is less than the set tolerance, proceed to step 6; otherwise, let... Return to step 2;
[0277] Step 6: Perform global normalization on the joint probability tensor obtained in the last iteration to obtain the final joint probability tensor.
[0278] Step 10: Sum the joint probability tensor obtained in Step 10 along the complementary dimensions of each dimension to obtain the marginal association probability of each dimension, and generate the Top-K association hypothesis with the highest probability based on the marginal association probability of each dimension and the bundle search.
[0279] Specifically,
[0280] Step A: Calculate the marginal association probability vector for each dimension. Taking the second dimension as an example:
[0281]
[0282] in, This is the marginal association probability vector of the second dimension. For measurement combination The corresponding value in the joint probability tensor;
[0283] Step B: Initialize beam width ,in, For the first The number of measurements from each sensor;
[0284] Step C: Sort the measurements of the first dimension in descending order of marginal association probability, and take the top results from the sorted list. The first measure is selected as a candidate measure; then, the measures of the second dimension are sorted in descending order of marginal association probability, and the top measures are selected from the sorted results. If each measurement is used as a candidate measurement, then the following combinations are generated: Each partial allocation has a heuristic probability that is the product of the marginal correlation probabilities of the two measures in the partial allocation. The allocation with the highest heuristic probability is retained. Each part is allocated;
[0285] Step D: Initialize the dimension ;
[0286] Step E: Arrange the first [item] in descending order of marginal association probability. The measurements in each dimension are sorted, and the top-ranked measurements are selected from the sort results. One measurement was used as a candidate measurement;
[0287] Using the newly extracted front Individual measurements and currently retained The combination of parts allocation results in Each partial allocation (the heuristic probability of the combined partial allocation is the product of the marginal correlation probability of the newly extracted measurement and the heuristic probability of the original partial allocation), and then from the obtained In the partial allocation, the part with the highest heuristic probability is retained. Each part is allocated;
[0288] Step F: Determine the dimension Is it equal to 5?
[0289] like If it equals 5, then the final value will be retained. Each assignment is used as a Top-K association hypothesis;
[0290] like If it is not equal to 5, then let Return to step E.
[0291] Step 11: Build a tracker consisting of JPDA and Kalman filter, and obtain the target coordinates based on weighted least squares and Top-K correlation assumption. Input the target coordinates into the tracker to update the target position.
[0292] Implementation example:
[0293] This example presents a set of simulation experiments to verify the multi-platform tracking method based on augmented Lagrange and probabilistic relaxation assignment proposed in this invention. The results of this invention are compared with those of the traditional Lagrange relaxation algorithm without probabilistic relaxation (referred to as the RL algorithm). Simulation results considering missed detections, noise, and clutter interference in multi-target tracking scenarios are given. During the simulation, a target signal is generated by simulation and a certain amount of noise is added. The underwater sound speed is taken as... The platform is equipped with a 16-element orthogonal cross array for receiving signals, with an element spacing of [missing information]. azimuth resolution is Five platforms observed simultaneously at 1-second intervals, with the tracking process lasting 60 seconds. A certain amount of random clutter existed at each moment, following a Poisson distribution with an expected value of 0.3. Azimuth measurements contained measurement noise with a covariance of 0.1. Azimuth measurements were obtained using the following method: first, MUSIC beamforming was used to obtain the beam response at each moment; then, a peak search method was used to extract the azimuth. The azimuth measurement was input into a tracking framework consisting of an S-dimensional allocator, a weighted least squares locator, a JPDA tracker, and a Kalman filter. Track initiation was initiated if a target was detected in three consecutive frames out of every five frames. Figure 4 This invention demonstrates the tracking results after processing data using augmented Lagrange and probabilistic relaxation assignment algorithms. Figure 5 The tracking results are shown without data processing using augmented Lagrange and probabilistic relaxation assignment algorithms.
[0294] It can be seen that the tracking results without augmented Lagrange and probabilistic relaxation assignment algorithms, when facing target intersections, cause the tracker to regenerate the trajectory due to the existence of multiple assignments with similar costs, misjudging it as the generation of a new target. The algorithm proposed in this invention solves this problem well, and can correctly judge the number of targets when three targets intersect, and the tracking trajectory is also smoother. Figure 6 This is a comparison chart of the root mean square error of the two algorithms under different signal-to-noise ratios. Figure 7 A comparison chart of the generalized optimal submode allocation of the two algorithms under different signal-to-noise ratios shows that the performance metrics of both algorithms are... The interval shows a significant improvement, which closely matches the spatial convergence period of the three targets in the scenario design. During this period, the spatial distance between targets decreases significantly, and the angle measurements acquired by each platform form multiple hypotheses in the multidimensional association cost matrix, drastically increasing the ambiguity of the S-dimensional allocation problem. For the RL algorithm, when the cost matrix exhibits a multi-valley structure, the sub-gradient update is prone to getting trapped in sub-optimal solutions, leading to measurement cross-correlation. Incorrect allocation forces the fusion of measurements from different targets, producing false detections with large positional deviations. Since the fusion covariance is not adjusted in real time according to the association uncertainty, the association gate of JPDA cannot accommodate this deviation, thus triggering the initialization of a new track, manifested as a significant peak in the GOSPA curve. For the algorithm of this invention, when convergence leads to increased association ambiguity, the system actively expands the measurement noise covariance of the fused detection, which is equivalent to expanding the association threshold of the tracker. This ensures that even if the fused position has a deviation due to residual allocation error, the current track can still be included in the gate with a high probability to complete the update, thereby suppressing track splitting and track initiation problems during the convergence period. Compared with the RL algorithm, the algorithm of this invention has the following advantages: At that time, the average GOSPA decreased The position RMSE decreased ;exist At that time, it reduced GOSPA The position RMSE decreased GOSPA, as a comprehensive indicator that simultaneously penalizes positioning errors, missed detections, and false tracks, shows a much larger reduction in performance than RMSE. This indicates that the performance gain of ALM-PRA mainly comes from the improvement in track maintenance accuracy, rather than simply the improvement in position estimation accuracy.
[0295] The above examples of the present invention are merely illustrative of the computational model and process of the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is impossible to exhaustively list all possible implementations here. Any obvious variations or modifications derived from the technical solutions of the present invention are still within the scope of protection of the present invention.
Claims
1. A multi-platform tracking method based on augmented Lagrange multiplication and probabilistic relaxation assignment, characterized in that, The method specifically includes the following steps: Step 1: Obtain The measurement data received by the sensor platform at the current sampling time, of which the first sensor platform will receive the measurement data at the current sampling time. The number of measurements received by the sensor platform at the current sampling time is denoted as . , ; Then according to The measurement data received by each sensor platform at the current sampling time is used to construct... Dimensional cost tensor, and according to S-dimensional cost tensor construction problem: Step 2: Initialize the number of iterations ; Step 3: Perform Lagrangian relaxation on the constructed S-dimensional allocation problem, decomposing it into a two-dimensional allocation subproblem, and obtaining the result of the two-dimensional allocation subproblem in the th... The cost corresponding to the next iteration; Step 4: Update the Lagrange multiplier vectors using the augmented Lagrange duality ascent mechanism and penalty parameters to obtain the cost of the measurement combination that satisfies all dimensional constraints. ; Step 5: Calculate the total constraint violation based on the constraint violation of each measurement, and adaptively adjust the penalty parameters according to the shrinkage characteristics of the total constraint violation. Step Six: Based on the two-dimensional allocator problem in the... The cost corresponding to the next iteration and Calculate the normalized dual gap, and determine whether the convergence condition of the augmented Lagrange algorithm is met based on the normalized dual gap and the total constraint violation. The convergence condition of the augmented Lagrange algorithm is to satisfy condition (1) or condition (2): Condition (1) Normalized dual gap converges; Condition (2) The total constraint violation degree converges; If the convergence condition of the augmented Lagrange algorithm is met, then step seven is executed using the final Lagrange multiplier vectors of each sensor platform. If the convergence condition of the augmented Lagrange algorithm is not met, then reconstruct the algorithm. Dimensional cost tensor, let Return to step three; Step 7: Use the final Lagrange multiplier vectors from each sensor platform to construct the last... The cost tensor is decoupled and the reduced cost tensor is calculated. Step 8: Adaptively select temperature parameters based on the statistical distribution of the reduced cost tensor. Then based on temperature parameters Compute set The Gibbs weights of the measurement combinations corresponding to each element in the dataset; Step 9: Perform alternating normalization on the Gibbs weights dimension by dimension to obtain the joint probability tensor; Step 10: Sum the joint probability tensor obtained in Step 10 along the complementary dimensions of each dimension to obtain the marginal association probability of each dimension, and generate the Top-K association hypothesis with the highest probability based on the marginal association probability of each dimension and the bundle search. Step 11: Build a tracker consisting of JPDA and Kalman filter, and obtain the target coordinates based on weighted least squares method and Top-K correlation assumption. Input the target coordinates into the tracker to update the target position. Continuous tracking of the target is achieved by updating the target position at each sampling time.
2. The multi-platform tracking method based on augmented Lagrange multiplication and probabilistic relaxation assignment according to claim 1, characterized in that, The S-dimensional allocation problem is as follows: in, This indicates the number of measurements received by the first sensor platform at the current sampling time. This indicates the number of measurements received by the second sensor platform at the current sampling time. This indicates the number of measurements received by the third sensor platform at the current sampling time. Indicates the first The number of measurements received by each sensor platform at the current sampling time. Indicates the first The number of measurements received by each sensor platform at the current sampling time; This indicates the number of samples received by the first sensor platform at the current sampling time. The first measurement, the second sensor platform received the first measurement at the current sampling time. The first measurement, ..., the first The sensor platform receives the first data at the current sampling time. The measurement combination consisting of several measurements comes from the negative log-likelihood cost of the same objective; Assign elements to the binary matrix; 。 3. The multi-platform tracking method based on augmented Lagrange multiplication and probabilistic relaxation assignment according to claim 2, characterized in that, The specific process of step three is as follows: Using Lagrange multiplier vectors The S-th constraint set of the relaxation S-dimensional assignment problem is obtained. Dimensional allocator problem: in, For the first The S-dimensional Lagrange multiplier vector after the nth iteration update ; For the first The first measurement received by the sensor platform at the current sampling time is in the... The Lagrange multipliers updated in the next iteration; For the first The second measurement received by the sensor platform at the current sampling time is in the... The Lagrange multipliers updated in the next iteration; For the first The sensor platform receives the first data at the current sampling time. The measurement was taken at the first The Lagrange multipliers updated in the next iteration; Similarly, we continue to use the Lagrange multiplier vectors to successively relax the S-dimensional allocation problem. constraint set, first The constraint set, ..., the third constraint set, decompose the constructed S-dimensional allocation problem into the following two-dimensional allocation subproblems: in, The cost of the two-dimensional allocator problem; Constructing new binary correlation variables : And define a new cost tensor for: Will and Substituting this into the two-dimensional allocator problem, we obtain the simplified two-dimensional allocator problem: Solving the two-dimensional allocator problem yields relaxed solutions and their corresponding costs. .
4. The multi-platform tracking method based on augmented Lagrange multiplication and probabilistic relaxation assignment according to claim 3, characterized in that, The specific process of step four is as follows: Step 4. Initialization ; Step 4.2: Apply the relaxation solution to the first step. Constraint set determination The cost of the dimension allocation problem and the dimensionality allocation problem The sensor platform receives the first data at the current sampling time. The measurement was taken at the first Constraint violation degree of the next iteration ; Step 43: Based on the degree of constraint violation Update # A Lagrange multiplier vector of dimension; Step 44, Order Determine whether the condition is met. ; If satisfied The cost of obtaining a combination of measurements that satisfies all dimensional constraints is... Then proceed to step five; If not satisfied Then return to step four two.
5. The multi-platform tracking method based on augmented Lagrange multiplication and probabilistic relaxation assignment according to claim 4, characterized in that, The specific process of step four three is as follows: in, For the first The sensor platform receives the first data at the current sampling time. The measurement was taken at the first The Lagrange multipliers updated in the next iteration; For the first The sensor platform receives the first data at the current sampling time. The measurement was taken at the first The Lagrange multipliers updated in the next iteration; For the first The penalty parameters are updated in the next iteration.
6. The multi-platform tracking method based on augmented Lagrange multiplication and probabilistic relaxation assignment according to claim 5, characterized in that, The specific process of step five is as follows: in, It is the L2 norm. Indicates the first Total constraint violation rate in each iteration Indicates the first The sensor platform in the first The constraint violation degree vector of the next iteration ; Indicates the first The first measurement received by the sensor platform at the current sampling time is in the... Constraint violation degree in the next iteration; Indicates the first The second measurement received by the sensor platform at the current sampling time is in the... Constraint violation degree in the next iteration; Indicates the first The sensor platform receives the first data at the current sampling time. The measurement was taken at the first Constraint violation degree in the next iteration; Then according to Adjust penalty parameters: in, As a penalty factor, Indicates the first Total constraint violation rate in each iteration Indicates the upper limit of the penalty parameter. The shrinkage threshold, For the first The penalty parameters are updated in the next iteration.
7. A multi-platform tracking method based on augmented Lagrange multiplication and probabilistic relaxation assignment according to claim 6, characterized in that, The specific process of step eight is as follows: in, Let be the set of all finite and non-zero elements in the reduced cost tensor. For set The standard deviation of all elements in the sample. These are the initial temperature parameters; Then temperature parameters for: in, For the final selected temperature parameters, For set The minimum value among all elements in; According to temperature parameters Compute set The Gibbs weights of the measurement combinations corresponding to each element in the dataset; in, Represents a set The measurement combination corresponding to the elements in Gibbs weight.
8. The multi-platform tracking method based on augmented Lagrange multiplication and probabilistic relaxation assignment according to claim 7, characterized in that, The specific process of step nine is as follows: Step 1: Initialize the number of iterations Initialize the joint probability tensor , The Gibbs weight tensor constructed based on the results of step nine; Step 2: Initialize the dimension; Step 3, for the first Normalize each dimension: in, In order to be in The joint probability tensor obtained in the nth iteration is the th The first dimension One value; In order to be in The joint probability tensor obtained in the nth iteration is the th The first dimension One value; For the joint probability tensor, the first The total number of values corresponding to each dimension; In order to be in The joint probability tensor obtained in the nth iteration is the th The first dimension One value; Step 4: Determine if the dimensions are satisfied. ; If satisfied Then proceed to step 5; If not satisfied, then let Return to step 3; Step 5, determine the first The joint probability tensor obtained in the second iteration is compared to the first iteration. If the change in the joint probability tensor obtained in the next iteration is less than the set tolerance, then proceed to step 6; otherwise, let... Return to step 2; Step 6: Perform global normalization on the joint probability tensor obtained in the last iteration to obtain the final joint probability tensor.
9. A multi-platform tracking method based on augmented Lagrange multiplication and probabilistic relaxation assignment according to claim 8, characterized in that, The specific process of step ten is as follows: Step A: Calculate the marginal association probability vector for each dimension: in, For the first A marginal correlation probability vector with multiple dimensions. For measurement combination The corresponding value in the joint probability tensor; Step B: Initialize beam width ,in, For the first The number of measurements from each sensor; Step C: Sort the measurements of the first dimension in descending order of marginal association probability, and take the top results from the sorted list. The first measure is used as a candidate measure. Then, the measures of the second dimension are sorted in descending order of marginal association probability. The top measures are selected from the sorted results. If each measurement is used as a candidate measurement, then the following combinations are generated: Each partial allocation has a heuristic probability that is the product of the marginal correlation probabilities of the two measures in the partial allocation. The allocation with the highest heuristic probability is retained. Each part is allocated; Step D: Initialize the dimension ; Step E: Arrange the first [item] in descending order of marginal association probability. The measurements in each dimension are sorted, and the top-ranked measurements are selected from the sort results. One measurement was used as a candidate measurement; Using the newly extracted front Individual measurements and currently retained The combination of the parts is obtained Each part is allocated, and then from the obtained In the partial allocation, the part with the highest heuristic probability is retained. Each part is allocated; Step F: Determine the dimension Is it equal to ; like equal Then it will be the one that is ultimately retained. Each assignment is used as a Top-K association hypothesis; like Not equal to Then let Return to step E.
10. A multi-platform tracking method based on augmented Lagrange multiplication and probabilistic relaxation assignment according to claim 9, characterized in that, The method for calculating the normalized dual gap is as follows: in, As an intermediate variable, For the first The cost of a relaxed solution to the two-dimensional assignment problem in the next iteration; in, For the first The normalized dual gap calculated in the next iteration This indicates taking the absolute value.
