Improved Transform track initiation method

Abstract

The invention relates to an improved Transform track initiation method, and belongs to the technical field of radar multi-target detection and tracking. Comprising the following steps: firstly, carrying out data preprocessing on radar measurement data to obtain a possible track, and carrying out feature extraction on the possible track; secondly, forming a classifier by using a Transform module and a Bayesian module, and carrying out dichotomy discrimination on the possible track to obtain the probability that the possible track is a real track; and finally, according to the probability, through maximum probability priority output, a final track initial result is obtained. According to the method, a binary classification network is constructed by adopting a Transform module and a Bayesian module, true and false track judgment is carried out on the possible track obtained after preprocessing, and the maximum probability priority output is carried out, so that effective starting of the true target track is finally realized, and the accuracy and the false alarm rate of track starting are improved.

Description

Technical Field

[0001] The invention relates to an improved Transformer track initiation method, and belongs to the technical field of radar multi-target detection and tracking. Background Art

[0002] The complex ocean environment has a significant impact on the performance of radar multi-target detection and tracking at sea. As the primary problem that needs to be solved in radar multi-target tracking, the track initiation method directly determines the accuracy of the subsequent target tracking process. In the radar multi-target tracking system, the track initiation method is first used to quickly and efficiently establish the target track to reduce the pressure of subsequent data processing, effectively eliminate clutter, and reduce the computational load of the filtering process. At the same time, the track initiation method helps to detect new track targets while performing effective track association, so as to minimize the problem of track missed detection.

[0003] The track initiation method refers to the process of determining the target track before the radar multi-target detection and tracking system enters a stable target tracking state. The classic track initiation methods are mainly divided into two categories: sequential processing methods and batch processing methods. Sequential processing methods include intuitive methods and logical methods. This method has a small amount of calculation and is easy to implement in engineering. It is mainly suitable for the track initiation of sparse echoes under weak clutter background. In a strong clutter environment, a batch processing method can be used. This method mainly includes the Hough transform method and its improved method, which can effectively reduce the false alarm probability. However, as the clutter density increases, the batch processing method will also have problems such as excessive calculation and poor real-time performance. Summary of the invention

[0004] In view of the deficiencies in the above-mentioned prior art, the present invention proposes an improved Transformer track initiation method, which adopts Transformer and Bayesian modules to construct a binary classification network, makes true and false track judgments on possible tracks obtained after preprocessing, and outputs them with maximum probability priority, thereby finally achieving effective initiation of the real target track, which is beneficial to improving the accuracy and false alarm rate of track initiation.

[0005] An improved Transformer track initiation method of the present invention is characterized in that it comprises the following steps:

[0006] Step 1: Preprocess the radar measurement data to obtain possible tracks and extract features of the possible tracks;

[0007] Step 2: Use the Transformer module and the Bayesian module to construct a classifier, perform binary classification on the possible tracks, and obtain the probability that the possible track is the real track;

[0008] Step 3: According to the probability size, the maximum probability is given priority to output to obtain the final track starting result.

[0009] Preferably, the specific steps of data preprocessing in step 1 are as follows:

[0010] According to the radar track information obtained from each radar scan, if the following conditions are met, they are considered to constitute a possible track;

[0011] Assume r i The position observations obtained by l consecutive scans, where i represents the number of consecutive scans, i = 1, 2, ..., l. If m observations satisfy the following three conditions in l scans, a possible track should be started:

[0012] 1) The average speed v between two adjacent scanning points of the target i , and the minimum speed V min and maximum value V max The following should be met:

[0013]

[0014] Among them, r i+1 -r i is the position interval obtained from two adjacent scans; t i+1 -t i is the time interval between two adjacent scans.

[0015] 2) The absolute value a of the acceleration of the target during three consecutive scans i , and the maximum acceleration a max The following should be met:

[0016]

[0017] 3) If a maneuvering target appears in the track association after scanning, in order to reduce the loss of the target or the formation of a false track, the velocity angle of the target is scanned three times in a row Angle with maximum speed The following are satisfied:

[0018]

[0019] V min 、V max 、a max , are the preset threshold values ​​of various parameters. Since the motion characteristic parameters of some false tracks cannot meet the threshold values ​​of various parameters preset in data preprocessing, some false tracks will be excluded in advance.

[0020] The specific steps of feature extraction in step 1 are as follows:

[0021] Select the 5 closest measurement points around the target track point, and fill in zeros if there are less than 5 points;

[0022] The feature f corresponding to the track point obtained by the i-th scan i Expressed as

[0023]

[0024] Among them, v i Represents the speed characteristic information corresponding to the track point obtained by the i-th scan; Represents the information of the yaw angle characteristics corresponding to the track point obtained by the i-th scan; r i1 、r i2 、r i3 、r i4 、r i5 They represent the distance information between the target track point (the i-th scan) and the five nearest measurement points (satisfying r i1 <r i2 <r i3 <r i4 <r i5 );θ i1 ,θ i2 ,θ i3 ,θ i4 ,θ i5 Respectively represent the angle information between the target track point (the i-th scan) and the five nearest measurement points;

[0025] Each possible track includes l-1 track points. If a track point is missing, it is necessary to supplement the feature information corresponding to the missing track point information according to the time. The supplementary information takes the average information of two adjacent points. The data feature set of each possible track is expressed as

[0026] F j =f1,f2,…f l-1 , j∈1,2,…,N (5)

[0027] Generate a feature set F=F corresponding to all possible track data sets including various motion features through feature extraction 1 , F 2 , …, F N .

[0028] Preferably, the feature extraction further comprises the following steps:

[0029] Label all tracks, and the labeled feature set can be used as the training sample set. The feature set of the real track is recorded as F real , the feature set of false tracks is recorded as F fault, the training sample set has a total of N labeled tracks, expressed as

[0030] F train =F 1 ,y 1 F 2 ,y 2 … F N ,y N (6)

[0031] in

[0032]

[0033] The Transformer module in step 2 adopts the Transformer encoder structure, which is composed of L identical encoder layers connected in series. Each encoder layer mainly includes two sublayers, where the first sublayer is multi-head self-attention and the second sublayer is a feedforward neural network. Each sublayer uses residual connection to fuse with the input data, and then inputs it to the next sublayer after normalization. The output dimension of each sublayer is designed to be d-dimensional.

[0034] The specific method of the Transformer module is as follows:

[0035] (1) Input sequence enters the MHSA layer to generate a new vector

[0036] Use the query vector (Q), key vector (K) and value vector (V) to assign attention weights to the input information, that is, decide which part of the input needs to be paid attention to and allocate limited information processing resources to the important part;

[0037] ① Input of MHSA layer

[0038] The input of MHSA is the feature data set of each starting track. j =f1,f2,…f l-1 , which can also be expressed as

[0039] F j =f i , i=1,2,…,l-1 (8)

[0040] ②Perform two linear transformations on the input information

[0041] The input vector passes through three weight matrices and the query matrix W q , key matrix W k , value matrix W v Perform the first linear transformation to obtain the query vector q i , key vector k i Sum value vector v i(where i∈1, 2, ..., l-1). Set Multi-Head to hHead, and then i Through the matrix W q1 , W q2 , …, W qh Perform the second linear transformation to obtain q i1 , Q i2 , ..., Q ih , and similarly for the key vector k i Through the matrix W k1 , W k2 , ..., W kh Perform the second linear transformation to get k i1 , k i2 , ..., k ih , for the value vector v i Through the matrix W v1 , W v2 , ..., W vh Perform the second linear transformation to get v i1 , v i2 , ..., v ih , the specific calculation formula is as follows:

[0042] Q ih =q i ·W qh =f i ·W q ·W qh (9)

[0043] k ih =k i ·W kh =f i ·W k ·W kh (10)

[0044] v ih =v i ·W vh =f i ·W v ·W vh (11)

[0045] Among them, W q , W k , W v are three trainable parameter matrices that can enhance the model's fitting ability. The initial value is a random initialization matrix that will be continuously updated during the training process; i = 1, 2, ..., l-1, h = 1, 2, ..., H, Q h =Q ih , i=1,2,…,l-1,K h =kih , i=1,2,…,l-1,V h =v ih , i=1,2,,l1,h all satisfy h∈1,2,…,H, then

[0046] Q h =F j ·W q ·W qh (12)

[0047] K h =F j ·W k ·W kh (13)

[0048] V h =F j ·W v ·W vh (14)

[0049] Among them, the obtained matrices Q, K and V can be directly used as the next encoder layer;

[0050] ③ Output of MHSA layer

[0051] The output attention matrix is ​​expressed as

[0052]

[0053] Among them, W O is the weight matrix of the given last linear projection layer; Q h K h T is the attention score, i.e., the matching degree; d n is the dimension of matrix Q, K, is the scale factor;

[0054] (2) The attention matrix Z output by the MHSA layer j With input F j Make a residual connection and normalize the result. The output is expressed as

[0055] Z j attention =LayerNorm(Z j +F j ) (16)

[0056] (3) The normalized result is placed in the FFN layer, i.e., the fully connected layer. The function of this layer is to perform a two-layer linear transformation on all position representations of the multi-head self-attention layer and activate it with the ReLU activation function;

[0057] Linear(Zj attention )=Z j attention ·w1+b1 (17)

[0058]

[0059] (4) The result Z output from step (3) j ffn Perform residual connection and normalization to get the output of the first encoder:

[0060] Z j =LayerNorm(Z j ffn +Z j attention ) (19)

[0061] (5) The output obtained from the first encoder is used as input and passed to the next encoder, which performs the same process as above until encoder L obtains the final output.

[0062] Preferably, the Bayesian module in step 2 is modeled using a Bayesian neural network to generate a probability distribution for the expected output result. The Bayesian neural network uses the output of the Transformer module as input, and uses the probability distribution instead of the fixed weight value in the traditional neural network to perform a binary classification task (real track or false track), and performs a secondary discrimination on the track start result of the data preprocessing;

[0063] (1) Perform Bayesian neural network modeling and set n weights, where the weight set W = w i , i = 1, 2, ..., n, where w i represents the i-th weight, each weight w i Subject to the mean μ i , with a variance of δ i Gaussian distribution, Gaussian distribution parameter θ i =(μ i , δ i ), and each weight is independent of each other. i and δ i It is the parameter that the network needs to train and update.

[0064] (2) Using a simple distribution qw i |θ i To approximate the output p W|Z j ,y j , calculate log qw respectively i |θ i 、logp w i 、log pyj |W,Z j .

[0065]

[0066] Among them, qw i |θ i Represents the distribution of weight parameters after given normal distribution parameters; pw i represents the prior distribution of weighted sample points, β is a random number between 0 and 1, and δ 01 ,δ 02 is the Gaussian distribution covariance set in advance; py j |W,Z j Represents given parameters W and Z j In the case of pred for y j The probability of y is the Gaussian distribution covariance set in advance;

[0067] (3) Calculate the loss function of the Bayesian neural network as follows:

[0068]

[0069] (4) Repeat the above steps (1) to (3) to update the parameters μi and δ i ;

[0070] (5) Output result py j |W,Z j , the same input needs to pass through the classifier multiple times to get multiple output results, and Z is calculated using the following formula j It is classified into whichever class it is most likely to be classified into:

[0071]

[0072] in, represents the label obtained after preprocessing of the jth starting track, m is 0 or 1. When m = 0, Represents Z j is a false track, when m=1, Represents Z j is the true track; Z j Represents Z j The probability of belonging to all categories, The feature vector Z representing the jth possible track j The categories obtained after the Bayesian neural network are real tracks or false tracks, with values ​​of 0 or 1 respectively.

[0073] Preferably, the step 3 specifically includes the following steps:

[0074] 1) The radar measurement data is acquired in real time, and then the data is preprocessed and screened using formulas (1) to (3), all false tracks are eliminated, and the possible track set F is obtained;

[0075] 2) Extract characteristic information of each possible track;

[0076] 3) The possible tracks are judged twice by the classifier in turn to obtain the probability value of the real track.

[0077] 4) Sort the probability values ​​corresponding to F in descending order;

[0078] 5) According to formula (8), if the maximum probability value is greater than 0.5, the corresponding track F′ j =f′ i , i = 1, 2, ..., l-1, j ∈ 1, 2, ..., each track point f′ in N′ i , i∈1, 2, …, l-1 all belong to this track, and move this track information to the new set F true If there is a track point f′ in the track i , i∈1, 2,…, l-1 appears in other tracks of F, directly delete these tracks from F;.

[0079] 6) Repeat steps 4) and 5) for the remaining tracks in F until all possible tracks have been screened;

[0080] 7) Output set F true , F true All the tracks in are the actual tracks finally obtained.

[0081] The present invention proposes an improved Transformer track initiation method, and introduces Transformer and Bayesian neural network into the track initiation method for the first time. By adopting a binary classification network constructed by Transformer and Bayesian modules, the true and false discrimination of possible tracks is realized, and the track can be judged more accurately, reducing the situation of misjudgment. At the same time, the strategy of maximum probability priority output further improves the accuracy of the starting result and improves the false alarm rate of track initiation. Therefore, the improved track initiation method proposed by the present invention can help to start tracking of marine targets more accurately, and is of great significance for the research on the reliability and effectiveness of the track initiation of unmanned targets, formation targets and group targets at sea, and provides an effective solution for marine target tracking and identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the conventional technology, the drawings required for use in the embodiments or the conventional technology descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0083] Figure 1 Schematic diagram of the structure of the classifier of the present invention. DETAILED DESCRIPTION

[0084] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0085] An improved Transformer track initiation method of this embodiment includes the following steps:

[0086] Step 1: Preprocess the radar measurement data to obtain possible tracks and extract features of the possible tracks;

[0087] Step 2: Use the Transformer module and the Bayesian module to construct a classifier, perform binary classification on the possible tracks, and obtain the probability that the possible track is the real track;

[0088] Step 3: According to the probability size, the maximum probability is given priority to output to obtain the final track starting result.

[0089] The specific steps of data preprocessing in step 1 are as follows:

[0090] According to the radar track information obtained from each radar scan, if the following conditions are met, they are considered to constitute a possible track;

[0091] Assume r i The position observations obtained by l consecutive scans, where i represents the number of consecutive scans, i = 1, 2, ..., l. If m observations satisfy the following three conditions in l scans, a possible track should be started:

[0092] 1) The average speed v between two adjacent scanning points of the target i , and the minimum speed V min and maximum value V max The following should be met:

[0093]

[0094] Among them, r i+1 -ri is the position interval obtained from two adjacent scans; t i+1 -t i is the time interval between two adjacent scans.

[0095] 2) The absolute value a of the acceleration of the target during three consecutive scans i , and the maximum acceleration a max The following should be met:

[0096]

[0097] 3) If a maneuvering target appears in the track association after scanning, in order to reduce the loss of the target or the formation of a false track, the velocity angle of the target is scanned three times in a row Angle with maximum speed The following are satisfied:

[0098]

[0099] V min 、V max 、a max , are the preset threshold values ​​of various parameters. Since the motion characteristic parameters of some false tracks cannot meet the threshold values ​​of various parameters preset in data preprocessing, some false tracks will be excluded in advance.

[0100] The specific steps of feature extraction in step 1 are as follows:

[0101] To extract the data features of the target track point, not only the characteristic information of the target itself but also the characteristic information of the target's surroundings should be considered. According to the nearest neighbor method, the information of the five closest measurement points around the target track point is selected, and zeros are added for those points less than five.

[0102] The feature f corresponding to the track point obtained by the i-th scan i Expressed as

[0103]

[0104] Among them, v i Represents the speed characteristic information corresponding to the track point obtained by the i-th scan; Represents the information of the yaw angle characteristics corresponding to the track point obtained by the i-th scan; r i1 、r i2 、r i3 、r i4 、r i5 They represent the distance information between the target track point (the i-th scan) and the five nearest measurement points (satisfying r i1 <r i2 <ri3 <r i4 <r i5 );θ i1 ,θ i2 ,θ i3 ,θ i4 ,θ i5 Respectively represent the angle information between the target track point (the i-th scan) and the five nearest measurement points;

[0105] Each possible track includes l-1 track points. If a track point is missing, it is necessary to supplement the feature information corresponding to the missing track point information according to the time. The supplementary information takes the average information of two adjacent points. The data feature set of each possible track is expressed as

[0106] F j =f1,f2,…f l-1 , j∈1,2,…,N (5)

[0107] Generate a feature set F=F corresponding to all possible track data sets including various motion features through feature extraction 1 , F 2 , …, F N .

[0108] Feature extraction also includes the following steps:

[0109] At present, the track set includes all possible target tracks and some false tracks, but these track feature sets are unlabeled and cannot be directly used as the input of the next classifier. Therefore, it is necessary to add labels to all tracks. The labeled feature set can be used as the training sample set. The feature set of the real track is recorded as F real , the feature set of false tracks is recorded as F fault , the training sample set has a total of N labeled tracks, expressed as

[0110] F train =F 1 ,y 1 F 2 ,y 2 … F N ,y N (6)

[0111] in

[0112]

[0113] The Transformer module in step 2 adopts the Transformer encoder structure, which is composed of L identical encoder layers connected in series. Each encoder layer mainly includes two sub-layers, where the first sub-layer is Multi-Head Self Attention (MHSA) and the second sub-layer is Feed Forward Network (FFN). Each sub-layer uses residual connection to fuse with the input data, which is then normalized and input to the next sub-layer. The output dimension of each sub-layer is designed to be d-dimensional.

[0114] The specific method of the Transformer module is as follows:

[0115] (1) Input sequence enters the MHSA layer to generate a new vector

[0116] Multi-head self-attention is not limited by feature space, and its modeling ability can meet the various complex relationships that may exist between data. The main purpose of using the attention mechanism is to use the query vector (Q), key vector (K) and value vector (V) to allocate attention weights to the input information, that is, to decide which part of the input needs to be paid attention to and allocate limited information processing resources to the important part;

[0117] ① Input of MHSA layer

[0118] The input of MHSA is the feature data set of each starting track. j =f1,f2,…f l-1 , which can also be expressed as

[0119] F j =f i , i=1,2,…,l-1 (8)

[0120] ②Perform two linear transformations on the input information

[0121] The input vector passes through three weight matrices and the query matrix W q , key matrix W k , value matrix W v Perform the first linear transformation to obtain the query vector Q i , key vector k i Sum value vector v i (where i∈1, 2, ..., l-1). Set Multi-Head to hHead, and then i Through the matrix W q1 , W q2 , …, W qh Perform the second linear transformation to obtain q i1 ,q i2, ..., q ih , and similarly for the key vector k i Through the matrix W k1 , W k2 , ..., W kh Perform the second linear transformation to get k i1 , k i2 , ..., k ih , for the value vector v i Through the matrix W v1 , W v2 , ..., W vh Perform the second linear transformation to get v i1 , v i2 , ..., v ih , the specific calculation formula is as follows:

[0122] Q ih =q i ·W qh =f i ·W q ·W qh (9)

[0123] k ih =k i ·W kh =f i ·W k ·W kh (10)

[0124] v ih =v i ·W vh =f i ·W v ·W vh (11)

[0125] Among them, W q , W k , W v are three trainable parameter matrices that can enhance the model's fitting ability. The initial value is a random initialization matrix that will be continuously updated during the training process; i = 1, 2, ..., l-1, h = 1, 2, ..., H, Q h =Q ih , i=1,2,…,l-1,K h =k ih , i=1,2,…,l-1,V h =v ih , i=1,2,…,l-1,h all satisfy h∈1,2,…,H, then

[0126] Q h =F j ·W q·W qh (12)

[0127] K h =F j ·W k ·W kh (13)

[0128] V h =F j ·W v ·W vh (14) The obtained matrices Q, K and V can be directly used as the next encoder layer.

[0129] ③ Output of MHSA layer

[0130] The output attention matrix is ​​expressed as

[0131]

[0132] Among them, W O is the weight matrix of the given last linear projection layer; Q h K h T is the attention score, i.e., the matching degree; d n is the dimension of matrix Q, K, is the scale factor.

[0133] (2) The attention matrix Z output by the MHSA layer j With input F j Make a residual connection and normalize the result. The output is expressed as

[0134] Z j attention =LayerNorm(Z j +F j ) (16)

[0135] (3) The normalized result is placed in the FFN layer, i.e., the fully connected layer. The function of this layer is to perform a two-layer linear transformation on all position representations of the multi-head self-attention layer and activate it with the ReLU activation function;

[0136] Linear(Z j attention )=Z j attention ·w1+b1 (17)

[0137]

[0138] (4) The result Z output from step (3) j ffnPerform residual connection and normalization to get the output of encoder 1:

[0139] Z j =LayerNorm(Z j ffn +Z j attention ) (19)

[0140] (5) The output obtained in encoder 1 is used as input and passed to the next encoder. The next encoder performs the same process as above until encoder L obtains the final output.

[0141] In step 2, the Bayesian module uses the Bayesian neural network to build a model, generate probability distribution for the expected output results, and enhance the robustness of the output results. The Bayesian neural network uses the output of the Transformer module as input, and uses probability distribution to replace the fixed weight value in the traditional neural network to perform a binary classification task (real track or false track). The track start result of data preprocessing is judged twice, thereby improving the accuracy of track start;

[0142] (1) Perform Bayesian neural network modeling and set n weights, where the weight set W = w i , i = 1, 2, ..., n, where w i represents the i-th weight, each weight w i Subject to the mean μ i , with a variance of δ i Gaussian distribution, Gaussian distribution parameter θ i =(μ i , δ i ), and each weight is independent of each other. i and δ i It is the parameter that the network needs to train and update.

[0143] (2) Using a simple distribution qw i |θ i To approximate the output p W|Z j ,y j , calculate log qw respectively i |θ i 、logp w i 、log py j |W,Z j .

[0144]

[0145] Among them, qw i |θ i Represents the distribution of weight parameters after given normal distribution parameters (because the weight sample point wi It is from qw i |θ i Sampling obtained); pw i represents the prior distribution of weighted sample points, β is a random number between 0 and 1, and δ 01 , 02 is the Gaussian distribution covariance set in advance; py j |W,Z j Represents given parameters W and Z j In the case of pred for y j The probability of y is the Gaussian distribution covariance set in advance;

[0146] (3) Calculate the loss function of the Bayesian neural network as follows:

[0147]

[0148] (4) Repeat the above steps (1) to (3) to update the parameter μ i and i ;

[0149] (5) Output result py j |W,Z j The same input needs to pass through the classifier multiple times to get multiple output results. Use the following formula to calculate Z j It is classified into whichever class it is most likely to be classified into:

[0150]

[0151] in, represents the label obtained after preprocessing of the jth starting track, m is 0 or 1. When m = 0, Represents Z j is a false track, when m=1, Represents Z j is the true track; Z j Represents Z j The probability of belonging to all categories, The feature vector Z representing the jth possible track j The categories obtained after the Bayesian neural network are real tracks or false tracks, with values ​​of 0 or 1 respectively.

[0152] The step 3 specifically includes the following steps:

[0153] 1) The radar measurement data is acquired in real time, and then the data is preprocessed and screened using formulas (1) to (3), all false tracks are eliminated, and the possible track set F is obtained;

[0154] 2) Extract characteristic information of each possible track;

[0155] 3) The possible tracks are judged twice by the classifier in turn to obtain the probability value of the real track.

[0156] 4) Sort the probability values ​​corresponding to F in descending order;

[0157] 5) According to formula (8), if the maximum probability value is greater than 0.5, the corresponding track F′ j =f′ i , i = 1, 2, ..., l-1, j ∈ 1, 2, ..., each track point f′ in N′ i , i∈1, 2, …, l-1 all belong to this track, and move this track information to the new set F true If there is a track point f′ in the track i , i∈1, 2,…, l-1 appears in other tracks of F, directly delete these tracks from F;.

[0158] 6) Repeat steps 4) and 5) for the remaining tracks in F until all possible tracks have been screened;

[0159] 7) Output set F true , F true All the tracks in are the actual tracks finally obtained.

[0160] Using the actual ship data set measured by marine radar, the proposed method is simulated and compared with the classic track initiation methods (logic method, intuitive method and Hough transform method) under the same clutter scene. min =0m / s, V max =20m / s, a max =0.125m / s 2 , Set the maximum longitude of the scene to 34, the minimum longitude to 28, the maximum latitude to 20.5, and the minimum latitude to 19.8. Sample the longitude and latitude in a uniform distribution to obtain the location coordinates of the clutter points. The number of clutter points ranges from 1000 to 10000, and each level increases by 1000 points, with a total of 10 clutter levels.

[0161] The results show that compared with the traditional track initiation method, this method is significantly superior in terms of track initiation accuracy and false alarm rate. Specifically, by adopting a binary classification network constructed by Transformer and Bayesian modules, the present invention can judge the track more accurately and reduce the possibility of misjudgment. At the same time, the strategy of maximum probability priority output further improves the accuracy of the initiation result. These results show that the improved track initiation method proposed in the present invention can more accurately initiate tracking of maritime targets, which is of great significance for the research on the reliability and effectiveness of track initiation of unmanned targets, formation targets and group targets at sea.

[0162] The above embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and all of them belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.

Claims

1. An improved Transformer track initiation method, characterized in that The following steps are involved: Step 1: Preprocess the radar measurement data to obtain possible tracks and extract features of the possible tracks; Step 2: Use the Transformer module and the Bayesian module to construct a classifier, perform binary classification on the possible tracks, and obtain the probability that the possible track is the real track; Step 3: According to the probability size, the maximum probability is given priority to output to obtain the final track starting result.

2. An improved Transformer track initiation method according to claim 1, characterized in that The specific steps of data preprocessing described in step 1 are as follows: According to the radar point track information obtained from each radar scan, if the following conditions are met, it is considered to constitute a possible track; Assume r i The position observations obtained by l consecutive scans, where i represents the number of consecutive scans, i = 1, 2, ..., l. If m observations satisfy the following three conditions in l scans, a possible track should be started: 1) The average speed v between two adjacent scanning points of the target i , and the minimum speed V min and maximum value V max The following should be met: Among them, r i+1 -r i is the position interval obtained from two adjacent scans; t i+1 t i is the time interval between two adjacent scans; 2) The absolute value a of the acceleration of the target during three consecutive scans i , and the maximum acceleration a max The following should be met: 3) If a maneuvering target appears in the track association after scanning, the velocity angle of the target for three consecutive scans Angle with maximum speed The following are satisfied: V min 、V max 、a max , are the preset threshold values ​​of various parameters.

3. An improved Transformer track initiation method according to claim 2, characterized in that The specific steps of feature extraction described in step 1 are as follows: Select the 5 closest measurement points around the target track point, and fill in zeros if there are less than 5 points; The feature f corresponding to the track point obtained by the i-th scan i It is expressed as: Among them, v i Represents the speed characteristic information corresponding to the track point obtained by the i-th scan; Represents the information of the yaw angle characteristics corresponding to the track point obtained by the i-th scan; r i1 、r i2 、r i3 、r i4 、r i5 Respectively represent the distance information between the target track point and the five nearest measurement points; θ i1 ,θ i2 ,θ i3 ,θ i4 ,θ i5 Respectively represent the angle information between the target track point and the five nearest measurement points; Each possible track includes l-1 track points. If a track point is missing, the feature information corresponding to the missing track point information needs to be supplemented according to the time. The supplementary information takes the average information of two adjacent points. The data feature set of each possible track is expressed as: F j =f1,f2,…f l-1 ,j∈1,2,…,N (5) Generate a feature set F=F corresponding to all possible track data sets including various motion features through feature extraction 1 ,F 2 ,…,F N .

4. An improved Transformer track initiation method according to claim 3, characterized in that The feature extraction in step 1 further includes the following steps: Label all tracks, and the labeled feature set can be used as the training sample set. The feature set of the real track is recorded as F real , the feature set of false tracks is recorded as F fault , the training sample set has a total of N labeled tracks, expressed as: F train =F 1 ,y 1 F 2 ,y 2 … F N ,y N (6) in:

5. An improved Transformer track initiation method according to claim 4, characterized in that The Transformer module described in step 2 adopts the Transformer encoder structure, which is composed of L identical encoder layers connected in series. Each encoder layer mainly includes two sub-layers, among which the first sub-layer is multi-head self-attention and the second sub-layer is a feedforward neural network. Each sub-layer uses residual connection to fuse with the input data, and then input to the next sub-layer after normalization. The output dimension of each sub-layer is designed to be d-dimensional.

6. An improved Transformer track initiation method according to claim 4, characterized in that The specific method of the Transformer module is as follows: (1) Input sequence enters the MHSA layer to generate a new vector Use the query vector Q, key vector K and value vector V to assign attention weights to the input information, that is, decide which part of the input needs to be paid attention to and allocate limited information processing resources to the important part; ① Input of MHSA layer The input of MHSA is the feature data set of each starting track. j =f1,f2,…f l1 , expressed as: F j =f i ,i=1,2,…,l-1 (8) ②Perform two linear transformations on the input information The input vector passes through three weight matrices and the query matrix W q , key matrix W k , value matrix W v Perform the first linear transformation to obtain the query vector q i , key vector k i Sum value vector v i , where i∈1,2,…,l-1; set Multi-Head to hHead, and then query vector q i Through the matrix W q1 , W q2 , ..., W qh Perform the second linear transformation to obtain q i1 ,q i2 , ..., q ih , and similarly for the key vector k i Through the matrix W k1 , W k2 , ..., W kh Perform the second linear transformation to get k i1 , k i2 , ..., k ih , for the value vector v i Through the matrix W v1 , W v2 , ..., W vh Perform the second linear transformation to get v i1 , v i2 , ..., v ih , the specific calculation formula is as follows: q ih =q i ·W qh =f i ·W q ·W qh (9) k ih =k i ·W kh =f i ·W k ·W kh (10) v ih =v i ·W vh =f i ·W v ·W vh (11) Among them, W q , W k , W v are three trainable parameter matrices, the initial values ​​are random initialization matrices, and they are continuously updated during the training process; i = 1, 2, ..., l1, h = 1, 2, ..., H, Q h =q ih ,i=1,2,…,l-1,K h =k ih ,i=1,2,…,l-1,V h =v ih ,i=1,2,…,l-1,h all satisfy h∈1,2,…,H, then: Q h =F j ·W q ·W qh (12) K h =F j ·W k ·W kh (13) V h =F j ·W v ·W· vh (14) Among them, the obtained matrices Q, K and V are directly used as the next encoder layer; ③ Output of MHSA layer The output attention matrix is ​​expressed as: Among them, W O is the weight matrix of the given last linear projection layer; Q h K h T is the attention score, i.e., the matching degree; d n is the dimension of matrix Q, K, is the scale factor; (2) The attention matrix Z output by the MHSA layer j With input F j Make a residual connection and normalize the result, the output is expressed as: Z j attention =LayerNorm(Z j +F j ) (16) (3) Put the normalized result into the FFN layer, i.e. the fully connected layer, and activate it with the ReLU activation function; Linear(Z j attention )=Z j attention w1+b1 (17) (4) The result Z output from step (3) j ffn Perform residual connection and normalization to get the output of the first encoder: WITH j =LayerNorm(Z j ffn +Z j attention ) (19) (5) The output obtained from the first encoder is used as input and passed to the next encoder, which performs the same process as above until encoder L obtains the final output.

7. An improved Transformer track initiation method according to claim 1, characterized in that The Bayesian module described in step 2 is modeled using a Bayesian neural network to generate a probability distribution for the expected output result. The Bayesian neural network uses the output of the Transformer module as input, and uses the probability distribution instead of the fixed weight value in the traditional neural network to perform a binary classification task and perform a secondary discrimination on the track start result of the data preprocessing; (1) Perform Bayesian neural network modeling and set n weights, where the weight set W = w i ,i=1,2,…,n, where w i represents the i-th weight, each weight w i Subject to the mean μ i , with a variance of δ i Gaussian distribution, Gaussian distribution parameter θ i =(μ i , i ), and each weight is independent of each other, μ i and x i It is the parameter that the network needs to train and update; (2) Using a simple distribution qw i |θ i To approximate the output p W|Z j ,y j , respectively calculate logq w i |θ i 、logp w i 、logpy j |W,Z j ; Among them, qw i |θ i Represents the distribution of weight parameters after given normal distribution parameters; pw i represents the prior distribution of weighted sample points, β is a random number in the interval 0,1, and δ 01 ,δ 02 is the Gaussian distribution covariance set in advance; py j |W,Z j Represents given parameters W and Z j In the case of pred for y j The probability of y is the Gaussian distribution covariance set in advance; (3) The loss function of the Bayesian neural network is calculated as: (4) Repeat the above steps (1) to (3) to update the parameter μ i and δ i ; (5) Output result py j |W,Z j , the same input needs to pass through the classifier multiple times to get multiple output results, and Z is calculated using the following formula j It is classified into whichever class it is most likely to be classified into: in, represents the label obtained after preprocessing of the jth starting track, m is 0 or 1. When m = 0, Represents Z j is a false track, when m=1, Represents Z j is the true track; Represents Z j The probability of belonging to all categories, The feature vector Z representing the jth possible track j The categories obtained after the Bayesian neural network are real tracks or false tracks, with values ​​of 0 or 1 respectively.

8. An improved Transformer track initiation method according to claim 6, characterized in that The step 3 specifically includes the following steps: 1) The radar measurement data is acquired in real time, and then the data is preprocessed and screened using formulas (1) to (3), all false tracks are eliminated, and the possible track set F is obtained; 2) Extract characteristic information of each possible track; 3) The possible tracks are sequentially identified by the classifier to obtain the probability value of the real track; 4) Sort the probability values ​​corresponding to F in descending order; 5) According to formula (8), if the maximum probability value is greater than 0.5, the corresponding track F′ j =f i ′,i=1,2,…,l-1,j∈1,2,…,each track point f in N i ′, i∈1,2,…,l-1 all belong to this track, and move this track information to the new set F true If there is a track point f in the track i ′, i∈1,2,…,l-1 appear in other tracks of F, directly delete these tracks from F; 6) Repeat steps 4) and 5) for the remaining tracks in F until all possible tracks have been screened; 7) Output set F true , F true All the tracks in are the actual tracks finally obtained.

Images