An Improved Transformer Track Initiation Method

Abstract

The present invention relates to an improved Transformer track initiation method, belonging to the technical field of radar multi-target detection and tracking. It includes: First, preprocess the radar measurement data to obtain possible tracks, and extract features from the possible tracks; Second, use a Transformer module and a Bayesian module to form a classifier to perform binary classification discrimination on the possible tracks to obtain the probability that the possible track is a true track; Finally, according to the probability size, output through the maximum probability first to obtain the final track initiation result. The present invention constructs a binary classification network using a Transformer and a Bayesian module to judge the true and false tracks of the possible tracks obtained after preprocessing, and outputs through the maximum probability first, so as to finally realize the effective initiation of the true target track, which is beneficial to improving the accuracy and false alarm rate of track initiation.

Description

Technical Field

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

[0002] Complex marine environments have a significant impact on the performance of radar in detecting and tracking multiple marine targets. As the primary problem to be solved in radar multi-target tracking, the track initiation method directly determines the accuracy of subsequent target tracking. In a radar multi-target tracking system, the track initiation method is first used to quickly and efficiently establish target tracks to reduce the subsequent data processing pressure, effectively eliminate clutter, and reduce the computational load during the filtering process. At the same time, the track initiation method helps to detect new track targets while performing effective track association to avoid track undetected problems as much as possible.

[0003] The track initiation method refers to the process of determining target tracks by a radar multi-target detection and tracking system before entering a stable target tracking state. Classic track initiation methods are mainly divided into two categories: sequential processing methods and batch processing methods. Sequential processing methods include intuitive methods, logical methods, etc. This method has a small computational amount and is easy to implement in engineering, and is mainly applicable to track initiation of sparse echoes in a weak clutter background. In a strong clutter environment, batch processing methods can be used. This method mainly includes the Hough transform method and its improved methods, etc., which can effectively reduce the false alarm probability. However, with the increase of clutter density, batch processing methods also have problems such as excessive computational amount 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. This method uses a Transformer and a Bayesian module to construct a binary classification network, performs true and false track judgments on the possible tracks obtained after preprocessing, and outputs the maximum probability first, so as to finally effectively initiate 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 by including the following steps:

[0006] Step 1: Perform data preprocessing on radar measurement data to obtain possible tracks, and extract features from the possible tracks;

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

[0008] Step 3: According to the probability magnitude, output by the maximum probability first, and obtain the final track initiation result.

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

[0010] Based on the radar point track information obtained from each radar scan, if the following conditions are met, it is considered that they form a possible track;

[0011] Assume r i is the position observation value obtained from l consecutive scans, where i represents the number of consecutive scans, i = 1, 2,..., l. If among the l scans, m observation values meet the following three conditions, it is determined that a possible track should be initiated:

[0012] 1) The average velocity v i between adjacent scan points of the target, and the minimum velocity V min and the maximum velocity V max should meet the following:

[0013]

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

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

[0016]

[0017] 3) If a maneuvering target appears in the track association after scanning, in order to reduce target loss or form false tracks, the included angle of the velocity of the target for three consecutive scans and the maximum included angle of velocity should meet the following:

[0018]

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

[0020] The specific steps of feature extraction in the said Step 1 are as follows:

[0021] Select the information of the 5 measurement points closest to the target track point around it. If there are less than 5 points, fill in zeros.

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

[0023]

[0024] where, v i represents the velocity feature information corresponding to the track point obtained by the i-th scan; represents the information of the yaw angle feature 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 (the i-th scan) and the 5 measurement points closest to it (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 5 measurement points closest to it;

[0025] Each possible track includes l - 1 track points. If a certain track point is missing, the feature information corresponding to the missing track point information in time needs to be supplemented according to time, and the supplementary information takes the average information of the adjacent two 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 the feature set F = F 1 , F 2 , …, F N .

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

[0029] Add labels to all tracks. The labeled feature set can be used as the training sample set. The feature set of the true track is denoted as F real , and the feature set of the false track is denoted as F fault, the training sample set has a total of N labeled tracks, denoted as

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

[0031] where

[0032]

[0033] The Transformer module in step 2 adopts a Transformer encoder structure, which is composed of L identical encoder layers connected in series. Each encoder layer mainly includes two sub-layers. Among them, the first sub-layer is the multi-head self-attention, and the second sub-layer is the feed-forward neural network; the residual connection is used to fuse the input data between each sub-layer, and after normalization, it is input to the next sub-layer. The output dimension of each sub-layer is designed to be d dimensions;

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

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

[0036] The query vector (Q), key vector (K), and value vector (V) are used to allocate the attention weights of the input information, that is, to determine which part of the input needs to be focused on, and to allocate the limited information processing resources to the important part;

[0037] ① Input of the MHSA layer

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

[0039] F j = f i , i = 1, 2, …, l - 1 (8)

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

[0041] The input vector passes through three weight matrices, the query matrix W q , the key matrix W k , and the value matrix W v to perform the first linear transformation to obtain the query vector q i , the key vector k i and the value vector v i(where \(i\in1, 2, \ldots, l - 1\)). Set Multi - Head to \(hHead\), and then for the query vector \(Q\) i Through the matrix \(W\) q1 , \(W\) q2 , \(\ldots\), \(W\) qh Perform a second - order linear transformation to obtain \(q\) i1 , \(Q\) i2 , \(\ldots\), \(Q\) ih , Similarly, for the key vector \(k\) i Through the matrix \(W\) k1 , \(W\) k2 , \(\ldots\), \(W\) kh Perform a second - order linear transformation to obtain \(k\) i1 , \(k\) i2 , \(\ldots\), \(k\) ih , For the value vector \(v\) i Through the matrix \(W\) v1 , \(W\) v2 , \(\ldots\), \(W\) vh Perform a second - order linear transformation to obtain \(v\) i1 , \(v\) i2 , \(\ldots\), \(v\) ih , The specific calculation formula is as follows:

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

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

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

[0045] Where \(W\) q , \(W\) k , \(W\) v Are three trainable parameter matrices, which can enhance the fitting ability of the model. The initial values are randomly initialized matrices and will be continuously updated during training; \(i = 1, 2, \ldots, l - 1\), \(h = 1, 2, \ldots, H\), \(Q\) h = \(Q\) ih , \(i = 1, 2, \ldots, l - 1\), \(K\) h = \(k\)ih , i = 1, 2, …, l - 1, V h = v ih , i = 1, 2,, l1, h all satisfy h ∈ 1, 2, …, H, then there is

[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 in the next encoder layer;

[0050] ③ Output of the MHSA layer

[0051] The output attention matrix is expressed as

[0052]

[0053] Among them, W O is the weight matrix of the given final linear projection layer; Q h K h T is the attention score, that is, the matching degree; d n is the dimension of matrices Q and K, is the scaling factor;

[0054] (2) Residual connection is made between the attention matrix Z j output from the MHSA layer and the input F j , and the result is normalized. The obtained output is expressed as

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

[0056] (3) The normalized result is put into the FFN layer, that is, the fully connected layer. The role of this layer is to perform 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) Feed the result Z output from step (3) j ffn through residual connection and normalization to obtain the output of the first encoder:

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

[0061] (5) Use the output obtained from the first encoder as the input and feed it into the next encoder. The next encoder performs the same processing as above until encoder L to obtain the final output.

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

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

[0064] (2) Use a simple distribution q w i |θ i to approximate the output p W|Z j , y j , and calculate log q w i |θ i , logp w i , log p yj |W, Z j 。

[0065]

[0066] Among them, q w i |θ i represents the distribution of the weight parameter after given the normal distribution parameters; p w i represents the prior distribution of the weight sample points, β is a random number in the interval of 0, 1, δ 01 and δ 02 are the covariance of the pre-set Gaussian distribution; p y j |W, Z j represents the network output y j under the condition of given parameters W and Z pred is y j probability, y is the covariance of the pre-set Gaussian distribution;

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

[0068]

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

[0070] (5) Output the result p y j |W, Z j , the same input needs to pass through the classifier multiple times to obtain multiple output results, and Z is calculated using the following formula j Whichever class has the highest probability of being classified into is the class it is assigned to:

[0071]

[0072] Among them, represents the label obtained after preprocessing the j-th starting track, m takes the value of 0 or 1, when m = 0, represents Z j is a false track, when m = 1, represents Z j is a true track; Z j represents Z j probability of belonging to all classes, represents the feature vector Z of the j-th possible track j obtained after passing through the Bayesian neural network, which is a true track or a false track, and takes the values of 0 or 1 respectively.

[0073] Preferably, the specific steps of step 3 include the following steps:

[0074] 1) Radar measurement data obtained in real time, and then data preprocessing and screening are carried out using formulas (1) to (3) to exclude all false tracks and obtain a set of possible tracks F;

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

[0076] 3) Pass the possible tracks through the classifier in sequence for secondary discrimination to obtain the probability value that it is a 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, then the corresponding track F′ j = f′ i , and each track point f′ i , where i ∈ 1, 2,..., l - 1, j ∈ 1, 2,..., N′ belongs to this track, and move the information of this track to the new set F true for storage. If there is a track point f′ i in this track, where 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 are screened;

[0080] 7) Output the set F true , F true and all the tracks in it are the finally obtained real tracks.

[0081] The present invention proposes an improved Transformer track initiation method, which first introduces Transformer and Bayesian neural network into the track initiation method. By using 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 misjudgment situation. At the same time, the strategy of outputting the maximum probability first further improves the accuracy of the initiation result and improves the false alarm rate of track initiation. Therefore, the improved track initiation method proposed by the present invention can help to more accurately initiate the tracking of maritime targets, and has important significance for the research on the reliability and effectiveness of track initiation of maritime unmanned targets, formation targets and group targets, providing an effective solution for maritime target tracking and recognition. Description of the Drawings

[0082] To more clearly illustrate the technical solutions in the embodiments of the present application or in the prior art, the following will briefly introduce the accompanying drawings required for the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other accompanying drawings can be obtained based on these drawings.

[0083] Figure 1 It is a schematic structural diagram of the classifier of the present invention. Detailed implementation manners

[0084] In order to make the objectives, technical solutions and advantages of the present application more clear and understandable, the present application will be further described in detail below with reference to 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 in this embodiment includes the following steps:

[0086] Step 1: Perform data preprocessing on the radar measurement data to obtain possible tracks, and extract features from the possible tracks;

[0087] Step 2: Use a Transformer module and a Bayesian module to form a classifier to perform binary classification discrimination on the possible tracks to obtain the probability that the possible tracks are real tracks;

[0088] Step 3: According to the probability magnitude, through the maximum probability first output, obtain the final track initiation result.

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

[0090] According to the radar point trace information obtained from each radar scan, if the following conditions are met, it is considered that they form a possible track;

[0091] Assume r i is the position observation value obtained from l consecutive scans, where i represents the number of consecutive scans, i = 1, 2,..., l. If among the l scans, m observation values meet the following three conditions, it is determined that a possible track should be initiated:

[0092] 1) The average velocity v i between the target's adjacent two scan points, and the minimum velocity V min and the maximum velocity V max should satisfy the following:

[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 obtained from two adjacent scans.

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

[0096]

[0097] 3) If a maneuvering target appears in the track association after scanning, in order to reduce target loss or the formation of false tracks, the angle between the speeds of the target for three consecutive scans and the maximum speed angle should satisfy the following:

[0098]

[0099] V min 、V max 、a max 、 are the threshold values of the preset parameters. Since the motion characteristic parameters of some false tracks cannot meet the threshold values of the preset parameters 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] Extract the data features of the target track points. Not only the target's own feature information but also the feature information around the target needs to be considered. According to the nearest neighbor method, select the information of the 5 measurement points closest to the target track point around it. If there are less than 5 points, fill in zeros;

[0102] The feature f corresponding to the track point obtained from the i-th scan i is expressed as

[0103]

[0104] where v i represents the speed feature information corresponding to the track point obtained from the i-th scan; represents the information of the yaw angle feature corresponding to the track point obtained from the i-th scan; r i1 、r i2 、r i3 、r i4 、r i5 respectively represent the distance information between the target track point (the i-th scan) and the 5 measurement points closest to it (satisfying r i1 <r i2 <ri3 < r i4 < r i5 ) ; θ i1 , θ i2 , θ i3 , θ i4 , θ i5 respectively represent the angular information between the target track point (the i-th scan) and the 5 measurement points closest to it;

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

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

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

[0108] Feature extraction also includes the following steps:

[0109] The current track set includes all possible target tracks and some false tracks, but the feature sets of these tracks are unlabeled and cannot be directly used as the input of the next classifier. Therefore, it is now necessary to add labels to all tracks, and the labeled feature sets can be used as the training sample set. The feature set of the true track is denoted as F real , and the feature set of the false track is denoted as F fault , and 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] where

[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. Among them, the first sub-layer is the multi-head self-attention (MHSA), and the second sub-layer is the feed-forward neural network (FFN); the input data is fused with the residual connection between each sub-layer, and after normalization, it is 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) The input sequence enters the MHSA layer to generate a new vector

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

[0117] ① Input of the MHSA layer

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

[0119] F j = f i , i = 1, 2,..., l - 1 (8)

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

[0121] The input vector passes through three weight matrices, the query matrix W q , the key matrix W k , the value matrix W v to perform the first linear transformation to obtain the query vector Q i , the key vector k i and the value vector v i (where i ∈ 1, 2,..., l - 1). Set Multi-Head to hHead, and then perform a second linear transformation on the query vector Q i through the matrices W q1 , W q2 , ……, W qh to obtain q i1 , q i2, ……, q ih , Similarly, for the key vector k i Through the matrix W k1 , W k2 , ……, W kh Perform a second linear transformation to obtain k i1 , k i2 , ……, k ih , For the value vector v i Through the matrix W v1 , W v2 , ……, W vh Perform a second linear transformation to obtain 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, which can enhance the fitting ability of the model. The initial values are randomly initialized matrices and will be continuously updated during training; 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, and h all satisfy h ∈ 1, 2, …, H, then there is

[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)Among them, the obtained matrices Q, K, and V can be directly used as the next encoder layer.

[0129] ③ Output of the MHSA layer

[0130] The output attention matrix is expressed as

[0131]

[0132] Among them, W O is the weight matrix of the given final linear projection layer; Q h K h T is the attention score, that is, the matching degree; d n is the dimension of matrices Q and K, is the scaling factor.

[0133] (2) Connect the attention matrix Z j output from the MHSA layer with the input F j and perform a residual connection on the result and normalize the result. The obtained output is expressed as

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

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

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

[0137]

[0138] (4) The result Z j ffnPerform residual connection and normalization to obtain the output of Encoder 1:

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

[0140] (5) Take the output obtained from Encoder 1 as the input and pass it to the next encoder. The next encoder performs the same processing as above until Encoder L to obtain the final output.

[0141] In step 2, the Bayesian module is modeled using a Bayesian neural network to generate a probability distribution for the expected output result, enhancing the robustness of the output result. The Bayesian neural network takes the output of the Transformer module as the input, uses the probability distribution to replace the fixed weight values in the traditional neural network, and performs a binary classification task (true track or false track) to perform a secondary discrimination on the track initiation result of the data preprocessing, thereby improving the accuracy of track initiation;

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

[0143] (2) Use a simple distribution q w i |θ i to approximate the output p W|Z j , y j , and calculate log q w i |θ i , logp w i , log p y j |W, Z j .

[0144]

[0145] Among them, q w i |θ i represents the distribution of the weight parameters given the normal distribution parameters (because the weight sample points wi is sampled from q w i |θ i ; p w i represents the prior distribution of the weighted sample points, β is a random number in the interval [0, 1], and δ 01 、 02 is the covariance of the Gaussian distribution set in advance; p y j |W, Z j represents the probability that the network output y j is y pred under the condition of given parameters W and Z j , and δ y is the covariance of the Gaussian distribution set in advance;

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

[0147]

[0148] (4) Repeat the above operations (1) - (3) to update the parameters μ i and i ;

[0149] (5) Output the result p y j |W, Z j . The same input needs to pass through the classifier multiple times to obtain multiple output results. Use the following formula to calculate Z j to determine which class has the highest possibility of being classified, and classify it into that class:

[0150]

[0151] where represents the label obtained after preprocessing the j-th starting track, m takes the value of 0 or 1. When m = 0, represents that Z j is a false track. When m = 1, represents that Z j is a true track; Z j represents the probability that Z j belongs to all classes, represents the feature vector Z j of the j-th possible track after passing through the Bayesian neural network, which is a true track or a false track, and takes the values of 0 or 1 respectively.

[0152] The specific steps of step 3 include the following steps:

[0153] 1) Real-time obtain the radar measurement data, and then use formulas (1) - (3) for data preprocessing and screening to exclude all false tracks and obtain the set F of possible tracks;

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

[0155] 3) Pass the possible tracks through the classifier for secondary discrimination in sequence to obtain the probability values that they are real tracks.

[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, then the corresponding track F′ j = f′ i , each track point f′ i , i ∈ 1, 2, …, l - 1, j ∈ 1, 2, …, N′ in the track belongs to this track, and move the information of this track to the new set F true for storage. If there is a track point f′ i , i ∈ 1, 2, …, l - 1 in this track that appears in other tracks in F, directly delete these tracks from F;

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

[0159] 7) Output the set F true , F true in which all tracks are the finally obtained real tracks.

[0160] Using the measured ship data set of marine radar, the method proposed by the present invention is simulated and compared with the classical track initiation methods (logical method, intuitive method and Hough transform method) in the same clutter scenario. The threshold values V min = 0 m / s, V max = 20 m / s, a max = 0.125 m / s 2 , Set the maximum longitude of the scenario 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 respectively in a uniform distribution to obtain the position coordinates of clutter points. The number of clutter points is from 1000 to 10000, increasing by 1000 points for each level, with a total of 10 clutter levels.

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

[0162] The above embodiments only illustrate several implementation manners of the present application, and the description thereof is relatively specific and detailed, but it should not be construed as a limitation on the scope of the invention patent. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can be made, 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 appended claims.

Claims

1. An improved Transformer track initiation method, characterized in that It includes the following steps: Step 1: Perform data preprocessing on the radar measurement data to obtain possible tracks, and extract features from the possible tracks; Step 2: Use a Transformer module and a Bayesian module to form a classifier to perform binary classification discrimination on the possible tracks, and obtain the probability that the possible track is a true track; Step 3: According to the probability size, output through the maximum probability first to obtain the final track starting result; The specific steps of the 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 that a possible track is formed; Assume r i is the position observation value obtained from l consecutive scans, where i represents the number of consecutive scans, i = 1, 2, ……, l. If among the l scans, m observation values satisfy the following three conditions, then it is determined that a possible track should be initiated: 1) The average speed v between two adjacent scanned points of the target i , and the minimum speed V min and the maximum speed V max should satisfy the following: where r i+1 - r i is the position interval obtained from two adjacent scans; t i+1 - t i is the time interval obtained from two adjacent scans; 2) The absolute value a of the acceleration of the target's three consecutive scans i , and the maximum acceleration a max should satisfy the following: 3) If a maneuvering target appears in the track association after scanning, the speed angle of the target in three consecutive scans and the maximum speed angle satisfy the following: V min 、V max 、a max 、 are the threshold values of the preset parameters; The specific steps of the feature extraction described in Step 1 are as follows: Select the information of the 5 measurement points closest to the target track point around the target track point, and fill in zeros if there are less than 5 points; Feature f corresponding to the track point obtained in the i-th scan i It is expressed as: Among them, v i represents the velocity feature information corresponding to the track point obtained from the i-th scan; represents the information of the yaw angle feature corresponding to the track point obtained from 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 5 closest measurement points; θ i1 、θ i2 、θ i3 、θ i4 、θ i5 respectively represent the angle information between the target track point and the 5 closest measurement points; Each possible track includes l - 1 track points. If a certain track point is missing, the feature information corresponding to the missing track point information in time needs to be supplemented according to time, and the supplementary information takes the average information of the adjacent two 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 1 , F 2 , …, F N} corresponding to all possible track data sets including various motion features through feature extraction.

2. An improved Transformer track initiation method according to claim 1, characterized in that The feature extraction described in Step 1 also includes the following steps: Add labels to all tracks, and the feature set with labels can be used as the training sample set. The feature set of real tracks is denoted as F real , and the feature set of false tracks is denoted as F fault , and 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) Where:

3. An improved Transformer track initiation method according to claim 2, 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 them, the first sub - layer is the multi - head self - attention, and the second sub - layer is the feed - forward neural network; the input data is fused with the residual connection between each sub - layer, and after normalization processing, it is input to the next sub - layer. The output dimension of each sub - layer is designed as d dimensions.

4. An improved Transformer track initiation method according to claim 2, characterized in that The specific method of the Transformer module is as follows: (1) The input sequence enters the MHSA layer to generate a new vector Use the query vector Q, key vector K, and value vector V to allocate the attention weights of the input information, that is, determine which part of the input needs to be focused on, and allocate the limited information processing resources to the important part; ① The input of the MHSA layer The input of MHSA is each starting track feature data set, and each starting track F j ={f1, f2, … f l-1}, which is expressed as: F j = {f i , i = 1, 2, …, l - 1} (8) ② Perform two linear transformations on the input information The input vector passes through three weight matrices, the query matrix W q , the key matrix W k , and the value matrix W v to perform the first linear transformation to obtain the query vector q i , the key vector k i and the value vector v i , where i ∈ {1, 2, …, l - 1}; set Multi - Head to hHead, and then for the query vector q i pass through the matrices W q1 , W q2 , ……, W qh to perform the second linear transformation to obtain q i1 , q i2 , ……, q ih , similarly for the key vector k i pass through the matrices W k1 , W k2 , ……, W kh to perform the second linear transformation to obtain k i1 , k i2 , ……, k ih , for the value vector v i pass through the matrices W v1 , W v2 , ……, W vh to perform the second linear transformation to obtain v i1 , v i2 , ……, v ih , and 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, with initial values being randomly initialized matrices and being 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}, and for all h satisfying h ∈ {1, 2, …, H}, there is: 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 The output of the MHSA layer The output attention matrix is expressed as: Among them, W O is the weight matrix of the given final linear projection layer; Q h (K h ) T is the attention score, that is, the matching degree; d n is the dimension of matrices Q and K, is the scaling factor; (2) Residual connection is made between the attention matrix Z output by the MHSA layer j and the input F j and the result is normalized. The obtained output is expressed as: Z j attention = LayerNorm(Z j + F j )(16) (3) Put the normalized result into the FFN layer, that is, the fully - connected layer, and activate it with the ReLU activation function; Linear(Z j attention ) = Z j attention ·w1 + b1 (17) (4) Residual connection and normalization are performed on the result Z output in step (3) j ffn to obtain the output of the first encoder: Z j = LayerNorm(Z j ffn + Z j attention )(19) (5) Take the output obtained in the first encoder as the input and pass it to the next encoder. The next encoder performs the processing of steps (1) - (4) until encoder L to obtain the final output.

5. An improved Transformer track initiation method according to claim 1, characterized in that The Bayesian module described in Step 2 uses a Bayesian neural network for modeling to generate a probability distribution for the expected output result. The Bayesian neural network takes the output of the Transformer module as the input, uses the probability distribution to replace the fixed weight value in the traditional neural network, and performs a binary classification task to perform secondary discrimination on the track starting result of the data preprocessing; (1) Perform Bayesian neural network modeling, set n weights, and the set of weights W = {w i , i = 1, 2, …, n}, where w i represents the i-th weight, and each weight w i follows a Gaussian distribution with mean μ i and variance δ i . The Gaussian distribution parameter θ i = (μ i , δ i ), and the weights are independent of each other. μ i and δ i are the parameters that the network needs to train and update; (2) Use a simple distribution q(w i |θ i ) to approximate the output p(W|Z j ,y j ), and calculate log q(w i |θ i ), logp(w i ), and log p(y j |W,Z j ) respectively; where q(w i |θ i ) represents the distribution of the weight parameter given the normal distribution parameters; p(w i ) represents the prior distribution of the weight sample points, β is a random number in the interval [0, 1], δ 01 and δ 02 are the pre-set Gaussian distribution covariances; p(y j |W, Z j ) represents the probability that the network output y j is y pred given the parameters W and Z j , and δ y is the pre-set Gaussian distribution covariance; (3) Calculate the loss function of the Bayesian neural network as: (4) Repeat the above operations (1) to (3) to update the parameters μ i and δ i ; (5) Output result p(y j |W, Z j ), the same input needs to pass through the classifier multiple times to obtain multiple output results, and Z is calculated using the following formula: j Whichever class has the highest likelihood of being classified into is the class it is assigned to: Among them, represents the label obtained after preprocessing the j-th starting track. m takes the value of 0 or 1. When m = 0, represents Z j is a false track. When m = 1, represents Z j is a true track; represents the probability that Z j belongs to all categories, represents the category obtained after the j-th possible track's feature vector Z j passes through the Bayesian neural network, which is a true track or a false track, taking the values of 0 or 1 respectively.

6. An improved Transformer track initiation method according to claim 4, characterized in that The specific content of Step 3 includes the following steps: 1) Real - time obtain the radar measurement data, and then use formulas (1) - (3) for data preprocessing screening to exclude all false tracks and obtain the possible track set F; 2) Extract the feature information of each possible track; 3) Pass the possible tracks through the classifier for secondary discrimination in turn to obtain the probability value that it is a 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, then the corresponding track F′ j ={f i ′, i = 1, 2, …, l - 1}, and each track point f i ′, i ∈ {1, 2, …, l - 1} in j ∈ {1, 2, …, N′} belongs to this track, and move the track information to the new set F true for storage. If there exists a track point f i ′, i ∈ {1, 2, …, l - 1} in this track that appears in other tracks in F, directly delete these tracks from F; 6) Repeat steps 4) and 5) for the remaining tracks in F until all possible tracks are screened; 7) Output set F true , F true All the tracks in it are the finally obtained true tracks.

Images