Flight path association method based on neural network

Abstract

The invention discloses a track association method based on a neural network. The method comprises the following steps: 1, preprocessing track data; 2, constructing a network model based on an attention mechanism; 3, feature processing of the trajectory data; 4, establishing an association probability loss function; 5, training a network model based on the attention mechanism; and 6, utilizing the trained network model based on the attention mechanism to carry out association prediction on a subsequent track. The method is simple in steps and reasonable in design, the feature point descriptors of the flight path serve as initial features, the association degree score matrix is obtained through feature updating and point product operation of the network model based on the attention mechanism, the association degree score matrix is optimized through the Sinkhorn algorithm to obtain the association probability matrix, the model efficiency is improved, and the method is suitable for being popularized and applied. Therefore, output of the association probability matrix is completed according to the input track characteristics, and track association is determined according to the association probability matrix.

Description

Technical Field

[0001] The invention belongs to the technical field of track association, and in particular relates to a track association method based on a neural network. Background Art

[0002] In the distributed multi-sensor multi-target tracking fusion system, it is necessary to determine whether the target tracks reported by different sensors originate from the same target, that is, track correlation. Track correlation is the premise of target information fusion and one of the key issues that need to be solved in target information fusion. At the same time, with the rapid development of military technology, the use of distributed radar networking to coordinate target detection and monitor battlefield situation has become an inevitable trend, and multi-sensor track correlation is particularly important.

[0003] At present, the methods for track association are mainly based on probability statistics, fuzzy mathematics, artificial intelligence, etc. Among them, the method based on probability statistics establishes statistical assumptions, uses the difference in state estimation as a statistic, and sets a threshold to determine whether the tracks are associated. This method is simple but requires manual setting of parameters, and as the number of targets and sensor errors increase, the accuracy of association decreases, and the scalability is not strong. The method based on fuzzy mathematics selects or designs the association membership, calculates the association membership of the two tracks to determine whether the tracks are associated. This type of method can handle the existing ambiguity, but the parameter setting is complex, and in practical applications, a lot of adjustment work may be required, which is not conducive to engineering implementation. The method based on artificial intelligence is to use neural networks to automatically extract track features, constrain track features in high-dimensional space through loss functions, and realize the mapping from track data to association results, but the track feature extraction has low correlation and the model efficiency is low.

[0004] To solve the above problems, a track association method based on neural network is needed. The characteristic point descriptor of the track is used as the initial feature. The association score matrix is ​​obtained through feature update and dot multiplication operation of the network model based on the attention mechanism. The Sinkhorn algorithm is used to optimize the association score matrix to obtain the association probability matrix to improve the efficiency of the model. The output of the association probability matrix is ​​completed according to the input trajectory features, and then the track association is determined according to the association probability matrix. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a track association method based on a neural network in view of the deficiencies in the above-mentioned prior art. The method has simple steps and reasonable design. The characteristic point descriptor of the track trajectory is used as the initial feature, and the correlation score matrix is ​​obtained through feature update and dot multiplication operation of a network model based on an attention mechanism. The Sinkhorn algorithm is used to optimize the correlation score matrix to obtain a correlation probability matrix, thereby improving the model efficiency, thereby completing the output of the correlation probability matrix based on the input track features, and then determining the track association based on the correlation probability matrix.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is: a track association method based on a neural network, characterized in that the method comprises the following steps:

[0007] Step 1: Preprocessing of track data:

[0008] Step 101: in the process of two sensors performing historical detection on multiple targets respectively, feature point descriptors of m targets detected by sensor A and feature point descriptors of n targets detected by sensor B are obtained; wherein m and n are both positive integers;

[0009] Step 102: Mark the feature point descriptor of the i-th target detected by sensor A and the feature point descriptor of the j-th target detected by sensor B. If they are related, the element P in the true association probability matrix is i,j Assign a value of 1. If there is no association, the element P in the true association probability matrix i,j Assign a value of 0; where i and j are both positive integers;

[0010] Step 103, repeating steps 101 and 102 multiple times, recording multiple feature point descriptors and their corresponding true association probability matrices as a historical data training set;

[0011] Step 2: Construction of network model based on attention mechanism:

[0012] Step 201, using a computer to establish a network model based on an attention mechanism; wherein the network model based on the attention mechanism includes a first MLP network module, an attention mechanism module, a second MLP network module and a dot product module; the attention mechanism module is composed of 9 identical attention layers stacked together, each attention layer includes a self-attention layer and a mutual attention layer, and the attention mechanism module includes a first attention layer, a second attention layer, a third attention layer, a fourth attention layer, a fifth attention layer, a sixth attention layer, a seventh attention layer, an eighth attention layer and a ninth attention layer;

[0013] Step 202: Use a computer to construct the first MLP network module MLPenc ;

[0014] Step 203: construct a second MLP network module using a computer;

[0015] Step 3: Feature processing of trajectory data:

[0016] Step 301: Using a computer according to Get the initial features of the i-th target detected by sensor A Among them, MLP enc (p i ) represents the feature point p of the i-th target detected by sensor A i Enter the first MLP network structure MLP enc Output features after d i represents the feature point descriptor of the i-th target detected by sensor A; where 1≤i≤m;

[0017] Step 302: According to the method of step 301, the initial feature of the jth target detected by sensor B is recorded as Among them, 1≤j≤n;

[0018] Step 303: Use a computer to and After the first attention layer, the second attention layer, the third attention layer, the fourth attention layer, the fifth attention layer, the sixth attention layer, the seventh attention layer, the eighth attention layer and the ninth attention layer, the eighteenth updated feature vector of the i-th target detected by sensor A is obtained. and the eighteenth updated feature vector of the jth target detected by sensor B

[0019] Step 304: Use a computer to update the eighteenth feature vector of the i-th target detected by sensor A and the eighteenth updated feature vector of the jth target detected by sensor B After being processed by the second MLP network module, the final updated feature vector f of the i-th target detected by sensor A is obtained i A and the final updated feature vector f of the jth target detected by sensor B j B ;

[0020] Step 305: Use a computer to update the final feature vector f of the i-th target detected by sensor A i A and the final updated feature vector f of the jth target detected by sensor B j BAfter the dot multiplication operation is performed by the dot multiplication module, the similarity value S between the i-th target detected by sensor A and the j-th target detected by sensor B is obtained. i,j ;

[0021] Step 306: Using a computer, the similarity values ​​between the multiple targets detected by sensor A and the multiple targets detected by sensor B are combined into a correlation score matrix;

[0022] Step 4: Establishment of the association probability loss function:

[0023] Step 401: Using a computer to optimize the association score matrix using the Sinkhorn algorithm to obtain a predicted association probability matrix;

[0024] Step 402: Use a computer to obtain the row target column numbers corresponding to the element value of 1 from the real association probability matrix, record them as target row numbers and target column numbers, and record all target row numbers and target column numbers as a to-be-judged set M;

[0025] Step 403: Using a computer according to Get the associated probability loss function L loss ; where e is a natural constant; i′ and j′ are the row and column numbers in the set to be judged M, P i″,j′ Represents the element value corresponding to the set M to be judged in the predicted association probability matrix;

[0026] Step 5: Training of network model based on attention mechanism:

[0027] Step 501: The computer uses the Adam optimization algorithm to use the associated probability loss function L loss Iteratively optimize the network model based on the attention mechanism until all historical data training sets are trained and one iterative training is completed;

[0028] Step 502, repeating the iterative training of step 501 until the preset number of iterative training times is met, and obtaining a trained network model based on the attention mechanism;

[0029] Step 6: Use the trained attention-based network model to predict the subsequent tracks:

[0030] Step 601: in the process of subsequent detection of multiple targets by two sensors respectively, a subsequent feature point descriptor of H targets detected by sensor A and a subsequent feature point descriptor of J targets detected by sensor B are obtained;

[0031] Step 602: input the subsequent feature point descriptors of H targets detected by sensor A and the subsequent feature point descriptors of J targets detected by sensor B into the network model based on the attention mechanism, and output the subsequent correlation score matrix;

[0032] Step 603: Use the Sinkhorn algorithm to process the subsequent correlation score matrix to obtain a subsequent correlation probability matrix. The element values ​​in the subsequent correlation probability matrix that are greater than 0.5 correspond to the track correlation of the two targets.

[0033] The above-mentioned track association method based on neural network is characterized in that: step 101, the specific process is as follows:

[0034] Step 1011, in the process of the sensor performing historical detection on multiple targets, the horizontal and vertical coordinates of the i-th target detected by the sensor A at each measurement time are obtained, and the horizontal and vertical coordinates of the i-th target detected by the sensor A at each measurement time are marked as each track point of the i-th target;

[0035] Step 1012: Combine the track points of the i-th target detected by sensor A into the feature point p of the i-th target detected by sensor A. i ; Among them, p i The number of track points is k, where k is a positive integer;

[0036] Step 1013: Obtain the feature point descriptor d of the i-th target detected by sensor A i ; where d i The number of feature description vectors is k, d i Any feature description vector in (x i ,y i ) represents the horizontal and vertical coordinates of the i-th target detected by sensor A, represents the radial velocity of the i-th target detected by sensor A, (sizex i ,sizey i ) represents the horizontal projection length and the vertical projection length of the i-th target detected by sensor A;

[0037] Step 1014: According to the method from step 1011 to step 1012, the feature point descriptors of the m targets detected by sensor A and the feature point descriptors of the n targets detected by sensor B are obtained.

[0038] The above-mentioned track association method based on neural network is characterized in that: in step 202, the first MLP network module MLP encIt includes an input layer, a first hidden layer, a second hidden layer, a third hidden layer and an output layer. The number of neurons in the input layer is 2, the number of neurons in the first hidden layer is 32, the number of neurons in the second hidden layer is 64, the number of neurons in the third hidden layer is 128, and the number of neurons in the output layer is 4;

[0039] The second MLP network module in step 203 includes an input layer, a hidden layer and an output layer. The number of neurons in the input layer is 4, the number of neurons in the hidden layer is 32, and the number of neurons in the output layer is 4.

[0040] The above-mentioned track association method based on neural network is characterized in that: step 303, the specific process is as follows:

[0041] Step 3031: Input the self-attention layer in the first attention layer, after being processed by the self-attention layer in the first attention layer, output the first updated feature vector of the i-th target detected by sensor A

[0042] At the same time, Input the self-attention layer in the first attention layer, after being processed by the self-attention layer in the first attention layer, output the first updated feature vector of the jth target detected by sensor B

[0043] Step 3032: and Input the mutual attention layer in the first attention layer, after being processed by the mutual attention layer in the first attention layer, the second updated feature vector of the i-th target detected by sensor A is output and the second updated feature vector of the jth target detected by sensor B

[0044] Step 3033: According to the method described in steps 3031 and 3032, the second updated feature vector of the i-th target detected by sensor A is and the second updated feature vector of the jth target detected by sensor B After being processed by the subsequent second to eighth attention layers, the eighteenth updated feature vector of the i-th target detected by sensor A is obtained. and the eighteenth updated feature vector of the jth target detected by sensor B

[0045] Compared with the prior art, the present invention has the following advantages:

[0046] 1. The method of the present invention has simple steps and reasonable design, and solves the current problem of track association.

[0047] 2. In the process of two sensors performing historical detection on multiple targets respectively, the present invention obtains feature point descriptors and true association probability matrices by preprocessing the track data, and uses them as historical data training sets.

[0048] 3. The present invention uses 9 alternating self-attention layers and mutual attention layers to update the features of the feature point descriptors corresponding to the track data to obtain the final updated feature vector; the representation ability is enhanced by the self-attention layer, and the correlation between the two is enhanced by the mutual attention layer, which reduces the calculation amount and parameter amount of the model and improves the model efficiency; and the final updated feature vectors of the two targets are point-multiplied to obtain the similarity values ​​of the two targets, and then form a correlation score matrix, and the feature representation of the aircraft data in the case of insufficient information and view occlusion is enhanced by alternating the self-attention layer and the mutual attention layer, thereby improving the robustness and accuracy of subsequent associations.

[0049] 4. The present invention uses the Sinkhorn algorithm to optimize the association score matrix, so that the association probability relationship can be directly output from the input trajectory information without additional manual processing or heuristic methods, thereby improving the efficiency and accuracy of the method.

[0050] In summary, the method of the present invention has simple steps and reasonable design. The characteristic point descriptor of the track trajectory is used as the initial feature, and the correlation score matrix is ​​obtained through feature update and dot multiplication operation of the network model based on the attention mechanism. The Sinkhorn algorithm is used to optimize the correlation score matrix to obtain the correlation probability matrix, thereby improving the model efficiency, thereby completing the output of the correlation probability matrix according to the input trajectory features, and then determining the track association according to the correlation probability matrix.

[0051] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 The figure is a flowchart of the method of the present invention. DETAILED DESCRIPTION

[0053] like Figure 1 As shown, the present invention provides a track association method based on a neural network, comprising the following steps:

[0054] Step 1: Preprocessing of track data:

[0055] Step 101: in the process of two sensors performing historical detection on multiple targets respectively, feature point descriptors of m targets detected by sensor A and feature point descriptors of n targets detected by sensor B are obtained; wherein m and n are both positive integers;

[0056] Step 102: Mark the feature point descriptor of the i-th target detected by sensor A and the feature point descriptor of the j-th target detected by sensor B. If they are related, the element P in the true association probability matrix is i,j Assign a value of 1. If there is no association, the element P in the true association probability matrix i,j Assign a value of 0; where i and j are both positive integers;

[0057] Step 103, repeating steps 101 and 102 multiple times, recording multiple feature point descriptors and their corresponding true association probability matrices as a historical data training set;

[0058] Step 2: Construction of network model based on attention mechanism:

[0059] Step 201, using a computer to establish a network model based on an attention mechanism; wherein the network model based on the attention mechanism includes a first MLP network module, an attention mechanism module, a second MLP network module and a dot product module; the attention mechanism module is composed of 9 identical attention layers stacked together, each attention layer includes a self-attention layer and a mutual attention layer, and the attention mechanism module includes a first attention layer, a second attention layer, a third attention layer, a fourth attention layer, a fifth attention layer, a sixth attention layer, a seventh attention layer, an eighth attention layer and a ninth attention layer;

[0060] Step 202: Use a computer to construct the first MLP network module MLP enc ;

[0061] Step 203: construct a second MLP network module using a computer;

[0062] Step 3: Feature processing of trajectory data:

[0063] Step 301: Using a computer according to Get the initial features of the i-th target detected by sensor A Among them, MLP enc (p i ) represents the feature point p of the i-th target detected by sensor A i Enter the first MLP network structure MLP enc Output features after d i represents the feature point descriptor of the i-th target detected by sensor A; where 1≤i≤m;

[0064] Step 302: According to the method of step 301, the initial feature of the jth target detected by sensor B is recorded as Among them, 1≤j≤n;

[0065] Step 303: Use a computer to and After the first attention layer, the second attention layer, the third attention layer, the fourth attention layer, the fifth attention layer, the sixth attention layer, the seventh attention layer, the eighth attention layer and the ninth attention layer, the eighteenth updated feature vector of the i-th target detected by sensor A is obtained. and the eighteenth updated feature vector of the jth target detected by sensor B

[0066] Step 304: Use a computer to update the eighteenth feature vector of the i-th target detected by sensor A and the eighteenth updated feature vector of the jth target detected by sensor B After being processed by the second MLP network module, the final updated feature vector f of the i-th target detected by sensor A is obtained i A and the final updated feature vector f of the jth target detected by sensor B j B ;

[0067] Step 305: Use a computer to update the final feature vector f of the i-th target detected by sensor A i A and the final updated feature vector f of the jth target detected by sensor B j B After the dot multiplication operation is performed by the dot multiplication module, the similarity value S between the i-th target detected by sensor A and the j-th target detected by sensor B is obtained. i,j ;

[0068] Step 306: Using a computer, the similarity values ​​between the multiple targets detected by sensor A and the multiple targets detected by sensor B are combined into a correlation score matrix;

[0069] Step 4: Establishment of the association probability loss function:

[0070] Step 401: Using a computer to optimize the association score matrix using the Sinkhorn algorithm to obtain a predicted association probability matrix;

[0071] Step 402: Use a computer to obtain the row target column numbers corresponding to the element value of 1 from the real association probability matrix, record them as target row numbers and target column numbers, and record all target row numbers and target column numbers as a to-be-judged set M;

[0072] Step 403: Using a computer according to Get the associated probability loss function L loss; where e is a natural constant; i′ and j′ are the row and column numbers in the set to be judged M, P i″,j′ Represents the element value corresponding to the set M to be judged in the predicted association probability matrix;

[0073] Step 5: Training of network model based on attention mechanism:

[0074] Step 501: The computer uses the Adam optimization algorithm to use the associated probability loss function L loss Iteratively optimize the network model based on the attention mechanism until all historical data training sets are trained and one iterative training is completed;

[0075] Step 502, repeating the iterative training of step 501 until the preset number of iterative training times is met, and obtaining a trained network model based on the attention mechanism;

[0076] Step 6: Use the trained attention-based network model to predict the subsequent tracks:

[0077] Step 601: in the process of subsequent detection of multiple targets by two sensors respectively, a subsequent feature point descriptor of H targets detected by sensor A and a subsequent feature point descriptor of J targets detected by sensor B are obtained;

[0078] Step 602: input the subsequent feature point descriptors of H targets detected by sensor A and the subsequent feature point descriptors of J targets detected by sensor B into the network model based on the attention mechanism, and output the subsequent correlation score matrix;

[0079] Step 603: Use the Sinkhorn algorithm to process the subsequent correlation score matrix to obtain a subsequent correlation probability matrix. The element values ​​in the subsequent correlation probability matrix that are greater than 0.5 correspond to the track correlation of the two targets.

[0080] In this embodiment, step 101, the specific process is as follows:

[0081] Step 1011, in the process of the sensor performing historical detection on multiple targets, the horizontal and vertical coordinates of the i-th target detected by the sensor A at each measurement time are obtained, and the horizontal and vertical coordinates of the i-th target detected by the sensor A at each measurement time are marked as each track point of the i-th target;

[0082] Step 1012: Combine the track points of the i-th target detected by sensor A into the feature point p of the i-th target detected by sensor A. i ; Among them, p i The number of track points is k, where k is a positive integer;

[0083] Step 1013: Obtain the feature point descriptor d of the i-th target detected by sensor A i ; where d i The number of feature description vectors is k, d i Any feature description vector in (x i ,y i ) represents the horizontal and vertical coordinates of the i-th target detected by sensor A, represents the radial velocity of the i-th target detected by sensor A, (sizex i ,sizey i ) represents the horizontal projection length and the vertical projection length of the i-th target detected by sensor A;

[0084] Step 1014: According to the method from step 1011 to step 1012, the feature point descriptors of the m targets detected by sensor A and the feature point descriptors of the n targets detected by sensor B are obtained.

[0085] In this embodiment, the first MLP network module MLP in step 202 enc It includes an input layer, a first hidden layer, a second hidden layer, a third hidden layer and an output layer. The number of neurons in the input layer is 2, the number of neurons in the first hidden layer is 32, the number of neurons in the second hidden layer is 64, the number of neurons in the third hidden layer is 128, and the number of neurons in the output layer is 4;

[0086] The second MLP network module in step 203 includes an input layer, a hidden layer and an output layer. The number of neurons in the input layer is 4, the number of neurons in the hidden layer is 32, and the number of neurons in the output layer is 4.

[0087] In this embodiment, step 303, the specific process is as follows:

[0088] Step 3031: Input the self-attention layer in the first attention layer, after being processed by the self-attention layer in the first attention layer, output the first updated feature vector of the i-th target detected by sensor A

[0089] At the same time, Input the self-attention layer in the first attention layer, after being processed by the self-attention layer in the first attention layer, output the first updated feature vector of the jth target detected by sensor B

[0090] Step 3032: and Input the mutual attention layer in the first attention layer, after being processed by the mutual attention layer in the first attention layer, the second updated feature vector of the i-th target detected by sensor A is output and the second updated feature vector of the jth target detected by sensor B

[0091] Step 3033: According to the method described in steps 3031 and 3032, the second updated feature vector of the i-th target detected by sensor A is and the second updated feature vector of the jth target detected by sensor B After being processed by the subsequent second to eighth attention layers, the eighteenth updated feature vector of the i-th target detected by sensor A is obtained. and the eighteenth updated feature vector of the jth target detected by sensor B

[0092] In this embodiment, the evaluation of the association result adopts the accuracy and recall rate indicators. Among the m targets detected by sensor A, the association method completes m of them. p The association of targets, among which the correct association is m t Among the n targets detected by sensor B, the association method completes n of them p The association of targets, among which the correct association is n t ; then:

[0093] In the formula, P is the precision, R is the recall, and m is p =n p , m t =n t ; l is the target that sensor A and sensor B simultaneously view and match;

[0094] In this embodiment, in a specific experiment, when m=n, that is, each sensor can detect 13 targets, then l=m=n, k is equal to 4, 8 and 20, and the interval is five seconds, and the association results of the trajectory points in the complete common view scene are obtained as shown in Table 1 below.

[0095] Table 1. Correlation results of trajectory points in complete common view scenario

[0096] k 4 8 20 Precision P 94.6％ 97.6％ 99.0％ Recall R 94.0％ 96.9％ 98.9％

[0097] In this embodiment, in the specific experiment, when m≠n, that is, each sensor can detect no more than 13 targets, then l<13 , k is equal to 4, 8 and 20, and the interval is five seconds, and the association results of trajectory points in some common view scenarios are obtained as shown in Table 2 below.

[0098] Table 2. Correlation results of trajectory points in some common view scenarios

[0099] k 4 8 20 Precision P 83.7％ 93.0％ 96.5％ Recall R 89.1％ 95.9％ 99.1％

[0100] Therefore, the present invention is suitable for aircraft association prediction in both complete co-viewing scenarios and partial co-viewing scenarios.

[0101] In summary, the method of the present invention has simple steps and reasonable design. The characteristic point descriptor of the track trajectory is used as the initial feature, and the correlation score matrix is ​​obtained through feature update and dot multiplication operation of the network model based on the attention mechanism. The Sinkhorn algorithm is used to optimize the correlation score matrix to obtain the correlation probability matrix, thereby improving the model efficiency, thereby completing the output of the correlation probability matrix according to the input trajectory features, and then determining the track association according to the correlation probability matrix.

[0102] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any way. Any simple modification, change and equivalent structural change made to the above embodiment based on the technical essence of the present invention still falls within the protection scope of the technical solution of the present invention.

Claims

1. A track association method based on neural network, characterized in that: The method comprises the following steps: Step 1: Preprocessing of track data: Step 101: in the process of two sensors performing historical detection on multiple targets respectively, feature point descriptors of m targets detected by sensor A and feature point descriptors of n targets detected by sensor B are obtained; wherein m and n are both positive integers; Step 102: Mark the feature point descriptor of the i-th target detected by sensor A and the feature point descriptor of the j-th target detected by sensor B. If they are related, the element P in the true association probability matrix is i,j Assign a value of 1. If there is no association, the element P in the true association probability matrix i,j Assign a value of 0; where i and j are both positive integers; Step 103, repeating steps 101 and 102 multiple times, recording multiple feature point descriptors and their corresponding true association probability matrices as a historical data training set; Step 2: Construction of network model based on attention mechanism: Step 201, using a computer to establish a network model based on an attention mechanism; wherein the network model based on the attention mechanism includes a first MLP network module, an attention mechanism module, a second MLP network module and a dot product module; the attention mechanism module is composed of 9 identical attention layers stacked together, each attention layer includes a self-attention layer and a mutual attention layer, and the attention mechanism module includes a first attention layer, a second attention layer, a third attention layer, a fourth attention layer, a fifth attention layer, a sixth attention layer, a seventh attention layer, an eighth attention layer and a ninth attention layer; Step 202: Use a computer to construct the first MLP network module MLP enc ; Step 203: construct a second MLP network module using a computer; Step 3: Feature processing of trajectory data: Step 301: Using a computer according to Get the initial features of the i-th target detected by sensor A Among them, MLP enc (p i ) represents the feature point p of the i-th target detected by sensor A i Enter the first MLP network structure MLP enc Output features after d i represents the feature point descriptor of the i-th target detected by sensor A; where 1≤i≤m; Step 302: According to the method of step 301, the initial feature of the jth target detected by sensor B is recorded as Among them, 1≤j≤n; Step 303: Use a computer to and After the first attention layer, the second attention layer, the third attention layer, the fourth attention layer, the fifth attention layer, the sixth attention layer, the seventh attention layer, the eighth attention layer and the ninth attention layer, the eighteenth updated feature vector of the i-th target detected by sensor A is obtained. and the eighteenth updated feature vector of the jth target detected by sensor B Step 304: Use a computer to update the eighteenth feature vector of the i-th target detected by sensor A and the eighteenth updated feature vector of the jth target detected by sensor B After being processed by the second MLP network module, the final updated feature vector f of the i-th target detected by sensor A is obtained i A and the final updated feature vector of the jth target detected by sensor B Step 305: Use a computer to update the final feature vector f of the i-th target detected by sensor A i A and the final updated feature vector of the jth target detected by sensor B After the dot multiplication operation is performed by the dot multiplication module, the similarity value S between the i-th target detected by sensor A and the j-th target detected by sensor B is obtained. i,j ; Step 306: Using a computer, the similarity values ​​between the multiple targets detected by sensor A and the multiple targets detected by sensor B are combined into a correlation score matrix; Step 4: Establishment of the association probability loss function: Step 401: Using a computer to optimize the association score matrix using the Sinkhorn algorithm to obtain a predicted association probability matrix; Step 402: Use a computer to obtain the row target column numbers corresponding to the element value of 1 from the real association probability matrix, record them as target row numbers and target column numbers, and record all target row numbers and target column numbers as a to-be-judged set M; Step 403: Using a computer according to Get the associated probability loss function L loss ; where e is a natural constant; i′ and j′ are the row and column numbers in the set to be judged M, P i″,j′ Represents the element value corresponding to the set M to be judged in the predicted association probability matrix; Step 5: Training of network model based on attention mechanism: Step 501: The computer uses the Adam optimization algorithm to use the associated probability loss function L loss Iteratively optimize the network model based on the attention mechanism until all historical data training sets are trained and one iterative training is completed; Step 502, repeating the iterative training of step 501 until the preset number of iterative training times is met, and obtaining a trained network model based on the attention mechanism; Step 6: Use the trained attention-based network model to predict the subsequent tracks: Step 601: in the process of subsequent detection of multiple targets by two sensors respectively, a subsequent feature point descriptor of H targets detected by sensor A and a subsequent feature point descriptor of J targets detected by sensor B are obtained; Step 602: input the subsequent feature point descriptors of H targets detected by sensor A and the subsequent feature point descriptors of J targets detected by sensor B into the network model based on the attention mechanism, and output the subsequent correlation score matrix; Step 603: Use the Sinkhorn algorithm to process the subsequent correlation score matrix to obtain a subsequent correlation probability matrix. The element values ​​in the subsequent correlation probability matrix that are greater than 0.5 correspond to the track correlation of the two targets.

2. A track association method based on a neural network according to claim 1, characterized in that: Step 101: The specific process is as follows: Step 1011, in the process of the sensor performing historical detection on multiple targets, the horizontal and vertical coordinates of the i-th target detected by the sensor A at each measurement time are obtained, and the horizontal and vertical coordinates of the i-th target detected by the sensor A at each measurement time are marked as each track point of the i-th target; Step 1012: Combine the track points of the i-th target detected by sensor A into the feature point p of the i-th target detected by sensor A. i ; Among them, p i The number of track points is k, where k is a positive integer; Step 1013: Obtain the feature point descriptor d of the i-th target detected by sensor A i ; where d i The number of feature description vectors is k, d i Any feature description vector in (x i ,y i ) represents the horizontal and vertical coordinates of the i-th target detected by sensor A, represents the radial velocity of the i-th target detected by sensor A, (sizex i ,sizey i ) represents the horizontal projection length and the vertical projection length of the i-th target detected by sensor A; Step 1014: According to the method from step 1011 to step 1012, the feature point descriptors of the m targets detected by sensor A and the feature point descriptors of the n targets detected by sensor B are obtained.

3. A track association method based on a neural network according to claim 1, characterized in that: The first MLP network module MLP in step 202 enc It includes an input layer, a first hidden layer, a second hidden layer, a third hidden layer and an output layer. The number of neurons in the input layer is 2, the number of neurons in the first hidden layer is 32, the number of neurons in the second hidden layer is 64, the number of neurons in the third hidden layer is 128, and the number of neurons in the output layer is 4; The second MLP network module in step 203 includes an input layer, a hidden layer and an output layer. The number of neurons in the input layer is 4, the number of neurons in the hidden layer is 32, and the number of neurons in the output layer is 4.

4. A track association method based on a neural network according to claim 1, characterized in that: Step 303, the specific process is as follows: Step 3031: Input the self-attention layer in the first attention layer, after being processed by the self-attention layer in the first attention layer, output the first updated feature vector of the i-th target detected by sensor A At the same time, Input the self-attention layer in the first attention layer, after being processed by the self-attention layer in the first attention layer, output the first updated feature vector of the jth target detected by sensor B Step 3032: and Input the mutual attention layer in the first attention layer, after being processed by the mutual attention layer in the first attention layer, the second updated feature vector of the i-th target detected by sensor A is output and the second updated feature vector of the jth target detected by sensor B Step 3033: According to the method described in steps 3031 and 3032, the second updated feature vector of the i-th target detected by sensor A is and the second updated feature vector of the jth target detected by sensor B After being processed by the subsequent second to eighth attention layers, the eighteenth updated feature vector of the i-th target detected by sensor A is obtained. and the eighteenth updated feature vector of the jth target detected by sensor B

Images