A track association method based on track similarity segmentation

Abstract

The present invention discloses a track association method based on track similarity segmentation, which mainly solves the problems existing in existing track association algorithms in actual engineering applications, such as the difficulty in obtaining prior information, the difficulty in determining the association threshold, and the long association time. The method utilizes deep learning technology, starting from the association matrix, and using a neural network to deeply explore the tracks of multiple sensors. All track information of each sensor is integrated to obtain a track matrix. By performing similarity segmentation on the track matrix, the association relationship between different tracks is directly obtained, thereby completely avoiding the traversal calculation of a large number of tracks, reducing the manual modification and debugging of models and parameters, and greatly improving the association efficiency. The method is suitable for the track association between multiple sensors in areas with rigid or non-rigid deformation during the sensor observation process, and can achieve rapid and accurate association of multiple sensors and multiple targets in actual engineering applications.

Description

Technical Field

[0001] The present invention belongs to the field of information fusion processing, and involves the fusion and association of multi-source tracks. It is suitable for the track association between multiple sensors in areas with rigid or non-rigid deformation during the observation process, and can achieve fast and accurate association of multiple sensors and multiple targets. Background Art

[0002] Track correlation involves correlating a large number of target tracks acquired by different sensors to correctly match tracks of the same target from different sensors, eliminating redundant tracks and providing a technical foundation for subsequent tasks such as information fusion and situational awareness. Track correlation ensures that a target has only one track, guaranteeing track uniqueness and preventing situational confusion. It is a key technology and core link in multi-source information fusion. If track correlation is not performed or the track correlation effect is poor, a large number of redundant tracks will be generated in the current area, seriously affecting the accuracy of subsequent situational analysis. Existing track correlation algorithms can be divided into two categories: traditional algorithms and neural network intelligent algorithms. Both require the construction of complex track similarity measurement models and construct a correlation matrix by comparing the similarities of tracks between multiple sensors to complete the correlation task. However, the process of traversing and calculating track similarity for each sensor inevitably increases correlation time and affects the algorithm's correlation efficiency. When the number of targets in the association area is small, the association time required by the algorithm does not increase significantly. However, when the number of targets is large, the association time required by the algorithm will increase exponentially, making it difficult to use algorithms with lower complexity directly, and difficult to meet the real-time requirements of engineering applications with higher complexity. Furthermore, traditional association algorithms require researchers to derive them by applying theoretical results from multiple fields, such as fuzzy, evidence, probability statistics, and probability estimation, under different target motion models and measurement noise assumptions, combined with their own heuristic considerations. In practical applications, these algorithms suffer from unreasonable prior assumptions, inappropriate models, and difficulty determining thresholds. The models and threshold parameters used must be manually modified and debugged repeatedly using measured data, which consumes a significant amount of time and effort, and the final connection effect after debugging is difficult to achieve optimal. In summary, existing model methods cannot directly solve the track association problem in actual engineering projects, and there is still a large gap between them and actual needs.

[0003] In order to solve the above problems, this paper uses deep learning technology, starting from the correlation matrix, and uses neural networks to deeply mine multi-sensor tracks, fusing all the track information of each sensor to construct a track tensor. By performing similarity segmentation on the track tensor, the correlation relationship between different tracks is directly obtained, thereby completely avoiding the traversal calculation of a large number of tracks, reducing the repeated manual modification and debugging of models and parameters, greatly improving the correlation efficiency, and saving manpower and material resources. Summary of the Invention

[0004] The present invention discloses a track association method based on track similarity segmentation, which specifically includes the following technical measures: first, track merging is performed and a true association matrix is ​​constructed, the tracks of each sensor are normalized and then merged to obtain a track tensor, and the true association matrix is ​​constructed based on the corresponding positions of the associated track pairs in different sensors. Then, in order for the neural network to learn the association matrix from the track tensor, a suitable loss function is designed, which includes three parts: similarity loss, row constraint loss, and column constraint loss. The similarity loss ensures that the generated association matrix is ​​as similar as possible to the true association matrix, and the row and column constraint losses ensure that each track under different sensors has one and only one association relationship, avoiding duplicate associations and missed associations. Then, a track tensor similarity segmentation network is constructed, which includes a track feature extraction module and an association mapping module. The track feature extraction module is used to extract track features from the track tensor, and the association mapping module maps the track tensor into an association matrix based on the extracted track features. The difference between the association matrix generated by the mapping and the true association matrix is ​​compared, and the network parameters are updated using a stochastic gradient descent algorithm. The track tensor similarity segmentation network model is then trained using the training dataset, and the model parameters are saved after training. Finally, the network model is tested using the test dataset. If the correlation effect meets the requirements, the network model can be put into practical use. If the correlation effect does not meet the requirements, the model parameters are tuned and the network training and testing are repeated.

[0005] A track association method based on track similarity segmentation includes the following steps:

[0006] Step 1: The tracks of sensor A and sensor B are converted into standardized, dimensionless tracks through normalization. A track tensor is constructed based on the tracks from different sensors. The elements at each position in the track tensor correspond one-to-one with the corresponding normalized tracks to complete the track merging.

[0007] Step 2: Design the loss function. The loss function of this method consists of three parts: similarity loss, row constraint loss, and column constraint loss. Define the network to perform similarity segmentation on the track tensor and obtain the correlation matrix A. The actual correlation matrix is ​​A. r , where the matrix element 1 indicates association and 0 indicates no association;

[0008] Step 3: Construct a similarity segmentation network model. The network consists of two modules: a track feature extraction module and an association mapping module. The track feature extraction module consists of three parallel branches: a basic feature extraction layer, a local feature extraction layer, and a global feature extraction layer. The association mapping module consists of three parallel branches: a row association mapping layer, a column association mapping layer, and a global association mapping layer.

[0009] Step 4: Train and test the similarity segmentation network. The input of the network is the track tensor T, and the output is the association matrix A. The association result is obtained according to the distribution of the values ​​in the output association matrix A.

[0010] Furthermore, the step 1 specifically includes the following sub-steps:

[0011] Step 1.1: First, traverse all track points in each sensor and find the maximum and minimum values ​​of each dimension of the track coordinates in the current scene. Then, process each dimension separately. For each track point, subtract the corresponding minimum value from each dimension and divide it by the corresponding maximum minus minimum value to obtain a standardized, dimensionless normalized track.

[0012] Step 1.2, traverse sensor A and get the number of tracks N in sensor A A , traverse sensor B and get the number of tracks N in sensor B B , intercept the nearest T sampling points for all tracks and construct a shape with a size of [N A ,N B ,4T], where all elements of T0 are 0 and the shape is [N A ,N B ]’s initial incidence matrix A0, where all elements of A0 are 0;

[0013] Step 1.3: Arrange the tracks in sensor A and sensor B according to the track batch number from small to large, traverse sensor A, and obtain the i-th track in sensor A. X i The shape is [T, 2], traverse sensor B, and obtain the j-th track in sensor B X j The shape is [T,2], X i and X j Flatten each row of , and splice the flattened tracks in the column direction to get the fused track X i,j , X i,j The shape is [4T,1],

[0014]

[0015]

[0016]

[0017] Among them, || is the splicing operation, and To flatten the track, the track X will be fused later. i,j Put the initial track tensor T0 into the i-th row and j-th column to get the track tensor T;

[0018] Step 1.4, sort the tracks in sensor A and sensor B according to the track batch number from small to large, traverse sensor A, and obtain the i-th track in sensor A According to the associated annotations in the data set, the i Associated track X m , where m represents X m At the position of the track in sensor B arranged in ascending order according to the track batch number, set the element in the i-th row and m-th column of the initial correlation matrix A0 to 1, and obtain the real correlation matrix A r .

[0019] Furthermore, the step 2 specifically includes the following sub-steps:

[0020] Step 2.1: The total loss function is the weighted sum of similarity loss, row constraint loss, and column constraint loss. The total loss function is

[0021] L=λ1×L sim +λ2×L row +λ3×L col

[0022] Among them, λ1, λ2, λ3 are weighting coefficients;

[0023] Step 2.2: Similarity loss ensures that the generated correlation matrix is ​​as similar as possible to the true correlation matrix. In order to accurately measure the difference between elements in the matrix, the square error of each element in the correlation matrix obtained after similarity segmentation and the true correlation matrix is ​​used. The similarity loss is,

[0024]

[0025] where a i,j represents the element in row i and column j of A, Indicates A l The element in the i-th row and j-th column of , ||*||2 represents the 2-norm of the matrix;

[0026] Step 2.3, the row constraint loss ensures that each row can have at most one association pair, that is, the sum of the elements in any row of A is less than or equal to 1, to prevent repeated associations and missed associations. The row constraint loss is,

[0027]

[0028] Step 2.4, the column constraint loss ensures that each column can have at most one association pair, that is, the sum of the elements in any row of A is less than or equal to 1, to prevent repeated associations and missed associations. The column constraint loss is,

[0029]

[0030] Furthermore, the specific values ​​of λ1, λ2, and λ3 are 1, 1, and 1.

[0031] Furthermore, the step 3 specifically includes the following sub-steps:

[0032] Step 3.1: Construct the element feature extraction layer. The element feature extraction layer consists of two fully connected layers. First, the track tensor is flattened and its shape is [N A ,N B ,4T] changes to [N A ×N B ,4T], where each element corresponds to two tracks from sensor A and sensor B. Then, the element feature extraction layer composed of two fully connected layers is used to process each element in the flattened track tensor, that is, to process two tracks from different sensors separately, enlarge the dimension to 4 times the original dimension and then scale it back to the original dimension. The shape of the tensor changes to [N A ×N B ,4T]->[N A ×N B ,16T]->[N A ×N B ,4T], and finally restore the flattened track tensor to its original shape [N A ,N B ,4T], the GELU nonlinear activation function is connected between the two fully connected layers;

[0033] Step 3.2: Construct the local feature extraction layer. The local feature extraction layer first extracts the neighborhood information in the track tensor by a 3×3 convolution layer, fuses the neighbor track features, and sets its step size and padding to 1 to ensure that the matrix size does not change. Then, two 1×1 convolution layers are connected to amplify the number of channels of the 3×3 convolution layer output feature by 4 times and then scale it back to the original size, fusing the track information of different channels. The change of the tensor shape in the two 1×1 convolution layers is [N A ,N B ,4T]->[N A ,N B ,16T]->[N A ,N B ,4T], finally, layer normalization is added between the 3×3 convolution layer and the first 1×1 convolution layer to speed up the network convergence, and the GELU nonlinear activation function is added between the first 1×1 convolution layer and the second 1×1 convolution layer;

[0034] Step 3.3: Construct the global feature extraction layer. The global feature extraction layer consists of two global self-attention layers. First, the track tensor is flattened and its shape is [N A ,N B ,4T] changes to [NA ×N B ,4T], the flattened track tensor is regarded as a whole, and then the query matrix, key matrix and value matrix are calculated.

[0035] Q=XW q

[0036] K=XW k

[0037] V=XW v

[0038] Among them, Q, K, and V are query matrix, key matrix, and value matrix respectively, and W q 、W k 、W v are the query mapping matrix, key mapping matrix, and value mapping matrix respectively, and X is the flattened track vector. In each global self-attention layer, the shapes of Q, K, and V are [N A ×N B ,4T],W q 、W k 、W v The shape of is [4T,4T], and the shape of X is [N A ×N B ,4T], then calculate Q and K T The inner product of To make the training of the global self-attention layer more stable, we use SoftMax as the self-attention score.

[0039]

[0040] Where S is the self-attention score, and the shape is [N A ×N B ,4T], then calculate the inner product of the self-attention score and V as the output,

[0041]

[0042] Where H is the output. In each global attention layer, a fully connected layer is used to change the output dimension. The first layer amplifies the dimension to 4 times the original dimension, and the second layer scales it back to the original dimension. The GELU nonlinear activation function is connected between the two global self-attention layers.

[0043] Step 3.4: Construct row-association mapping layer, column-association mapping layer, and global-association mapping layer. The row-association mapping layer is a 1×3 convolutional layer with a stride of (1,1) and padding of (0,1). The number of input and output channels is the same. The column-association mapping layer is a 3×1 convolutional layer with a stride of (1,1) and padding of (1,0). The number of input and output channels is the same. The global-association mapping layer is a global self-attention layer. The track tensor is flattened as input and the output tensor is restored to its original shape. The input and output dimensions are the same.

[0044] Step 3.5: Merge the outputs of the row association mapping layer, column association mapping layer, and global association mapping layer according to the channel dimension, and perform association mapping merging through a convolutional layer with a convolution kernel of 1×1, a stride of 1, a padding of 0, and an output channel of 1 to obtain the association matrix A.

[0045] The proposed track association method, based on track similarity segmentation, automatically mines track features based on measured data, directly derives correlations between different tracks, and achieves adaptive track association. This method avoids the need to repeatedly calculate tracks from different sensors, reduces the need for manual model and parameter modification and debugging, significantly improves association efficiency, and saves manpower and resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is a flow chart of a track association method based on track similarity segmentation;

[0047] Figure 2 It is the structure diagram of the track similarity segmentation network;

[0048] Figure 3 It is the structure diagram of track tensor and incidence matrix. DETAILED DESCRIPTION

[0049] The present invention proposes a track association method based on track similarity segmentation, which includes the following steps:

[0050] Step 1: The tracks of sensor A and sensor B are converted into standardized, dimensionless tracks through normalization. A track tensor is constructed based on the tracks from different sensors. The elements at each position in the track tensor correspond one-to-one with the corresponding normalized tracks to complete the track merging.

[0051] Step 1.1: First, traverse all track points in each sensor and find the maximum and minimum values ​​of each dimension of the track coordinates in the current scene. Then, process each dimension separately. For each track point, subtract the corresponding minimum value from each dimension and divide it by the corresponding maximum minus minimum value to obtain a standardized, dimensionless normalized track.

[0052] Step 1.2: Traverse sensor A and get the number of tracks N in sensor AA , traverse sensor B and get the number of tracks N in sensor B B , intercept the nearest T sampling points for all tracks and construct a shape with a size of [N A ,N B ,4T], where all elements of T0 are 0 and the shape is [N A ,N B ]’s initial incidence matrix A0, where all elements of A0 are 0;

[0053] Step 1.3: Arrange the tracks in sensor A and sensor B according to the track batch number from small to large, traverse sensor A, and obtain the i-th track in sensor A. X i The shape is [T, 2], traverse sensor B, and obtain the j-th track in sensor B X j The shape is [T,2], X i and X j Flatten each row of , and splice the flattened tracks in the column direction to get the fused track X i,j , X i,j The shape is [4T,1],

[0054]

[0055]

[0056]

[0057] Among them, || is the splicing operation, and To flatten the track, the track X will be fused later. i,j Put the initial track tensor T0 into the i-th row and j-th column to get the track tensor T;

[0058] Step 1.4: Arrange the tracks in sensor A and sensor B according to the track batch number from small to large, traverse sensor A, and obtain the i-th track in sensor A. According to the associated annotations in the data set, the i Associated track X m , where m represents X m At the position of the track in sensor B arranged in ascending order according to the track batch number, set the element in the i-th row and m-th column of the initial correlation matrix A0 to 1, and obtain the real correlation matrix A r ;

[0059] Step 2: Design the loss function. The loss function of this method consists of three parts: similarity loss, row constraint loss, and column constraint loss. Define the network to perform similarity segmentation on the track tensor and obtain the correlation matrix A. The actual correlation matrix is ​​A. r , where the matrix element 1 indicates association and 0 indicates no association;

[0060] Step 2.1: The total loss function is the weighted sum of similarity loss, row constraint loss, and column constraint loss.

[0061] L=λ1×L sim +λ2×L row +λ3×L col (4)

[0062] Wherein λ1, λ2, λ3 are weighting coefficients, and in the present invention, λ1 = λ2 = λ3 = 1;

[0063] Step 2.2: Similarity loss ensures that the generated correlation matrix is ​​as similar as possible to the true correlation matrix. In order to accurately measure the difference between the elements in the matrix, the square error of each element in the correlation matrix obtained after similarity segmentation and the true correlation matrix is ​​used.

[0064]

[0065] Among them L sim is the similarity loss, a i,j represents the element in row i and column j of A, Indicates A r The element in the i-th row and j-th column of , ||*||2 represents the 2-norm of the matrix;

[0066] Step 2.3: Row constraint loss ensures that each row can have at most one association pair, that is, the sum of the elements in any row of A is less than or equal to 1, to prevent duplicate associations and missed associations.

[0067]

[0068] Among them L row represents the row constraint loss;

[0069] Step 2.4: Column constraint loss ensures that each column can have at most one association pair, that is, the sum of the elements in any row of A is less than or equal to 1, to prevent duplicate associations and missed associations.

[0070]

[0071] Among them L col represents the row constraint loss;

[0072] Step 3: Construct a similarity segmentation network model. The network consists of two modules: the track feature extraction module and the association mapping module. The track feature extraction module consists of three parallel branches: the basic feature extraction layer, the local feature extraction layer, and the global feature extraction layer. The association mapping module consists of three parallel branches: the row association mapping layer, the column association mapping layer, and the global association mapping layer.

[0073] Step 3.1: Construct the element feature extraction layer. The element feature extraction layer consists of two fully connected layers. First, the track tensor is flattened and its shape is [N A ,N B ,4T] changes to [N A ×N B ,4T], where each element corresponds to two tracks from sensor A and sensor B. Then, the element feature extraction layer composed of two fully connected layers is used to process each element in the flattened track tensor, that is, to process two tracks from different sensors separately, enlarge the dimension to 4 times the original dimension and then scale it back to the original dimension. The shape of the tensor changes to [N A ×N B ,4T]->[N A ×N B ,16T]->[N A ×N B ,4T], and finally restore the flattened track tensor to its original shape [N A ,N B ,4T], the GELU nonlinear activation function is connected between the two fully connected layers;

[0074]

[0075] Step 3.2: Construct the local feature extraction layer. The local feature extraction layer first extracts the neighborhood information in the track tensor by a 3×3 convolution layer, fuses the neighbor track features, and sets its step size and padding to 1 to ensure that the matrix size does not change. Then, two 1×1 convolution layers are connected to amplify the number of channels of the 3×3 convolution layer output feature by 4 times and then scale it back to the original size, fusing the track information of different channels. The change of the tensor shape in the two 1×1 convolution layers is [N A ,N B ,4T]->[N A ,N B ,16T]->[N A ,N B ,4T], finally, layer normalization is added between the 3×3 convolution layer and the first 1×1 convolution layer to speed up the network convergence, and the GELU nonlinear activation function is added between the first 1×1 convolution layer and the second 1×1 convolution layer;

[0076] Step 3.3: Construct the global feature extraction layer. The global feature extraction layer consists of two global self-attention layers. First, the track tensor is flattened and its shape is [N A ,N B ,4T] changes to [N A ×N B ,4T], the flattened track tensor is regarded as a whole, and then the query matrix, key matrix and value matrix are calculated.

[0077] Q=XW q (9)

[0078] K=XW k (10)

[0079] V=XW v (11)

[0080] Among them, Q, K, and V are query matrix, key matrix, and value matrix respectively, and W q 、W k 、W v are the query mapping matrix, key mapping matrix, and value mapping matrix respectively, and X is the flattened track vector. In each global self-attention layer, the shapes of Q, K, and V are [N A ×N B ,4T],W q 、W k 、W v The shape of is [4T,4T], and the shape of X is [N A ×N B ,4T], then calculate Q and K T The inner product of To make the training of the global self-attention layer more stable, we use SoftMax as the self-attention score.

[0081]

[0082] Where S is the self-attention score, and the shape is [N A ×N B ,4T], and finally calculate the inner product of the self-attention score and V as the output,

[0083]

[0084] Where H is the output. In each global attention layer, a fully connected layer is used to change the output dimension. The first layer amplifies the dimension to 4 times the original dimension, and the second layer scales it back to the original dimension. The GELU nonlinear activation function is connected between the two global self-attention layers.

[0085] Step 3.4: Construct row-association mapping layer, column-association mapping layer, and global-association mapping layer. The row-association mapping layer is a 1×3 convolutional layer with a stride of (1,1) and padding of (0,1). The number of input and output channels is the same. The column-association mapping layer is a 3×1 convolutional layer with a stride of (1,1) and padding of (1,0). The number of input and output channels is the same. The global-association mapping layer is a global self-attention layer. The track tensor is flattened as input and the output tensor is restored to its original shape. The input and output dimensions are the same.

[0086] Step 3.5: Merge the outputs of the row association map layer, column association map layer, and global association map layer according to the channel dimension, and perform association mapping merging through a convolutional layer with a convolution kernel of 1×1, a stride of 1, a padding of 0, and an output channel of 1 to obtain the association matrix A.

[0087] Step 4: Train and test the similarity segmentation network. The input of the network is the track tensor T, and the output is the association matrix A. The association result is obtained based on the distribution of the values ​​in the output association matrix A.

Claims

1. A track association method based on track similarity segmentation, characterized in that: The following steps are involved: Step 1: The tracks of sensor A and sensor B are converted into standardized, dimensionless tracks through normalization. A track tensor is constructed based on the tracks from different sensors. The elements at each position in the track tensor correspond one-to-one with the corresponding normalized tracks to complete the track merging. Step 2: Design the loss function. The loss function of this method consists of three parts: similarity loss, row constraint loss, and column constraint loss. The network is defined to perform similarity segmentation on the track tensor and obtain the correlation matrix as follows: , the true incidence matrix is , where the matrix element 1 indicates association and 0 indicates no association; Step 3: Construct a similarity segmentation network model. The network consists of two modules: the track feature extraction module and the association mapping module. The track feature extraction module consists of three parallel branches: the element feature extraction layer, the local feature extraction layer, and the global feature extraction layer. The association mapping module consists of three parallel branches: the row association mapping layer, the column association mapping layer, and the global association mapping layer. Step 3.1: Construct the element feature extraction layer. The element feature extraction layer consists of two fully connected layers. First, the track tensor is flattened and its shape is Changes to Then, the element feature extraction layer composed of two fully connected layers is used to process each element in the flattened track tensor, enlarging the dimension to 4 times the original dimension and then scaling it back to the original dimension. The shape of the tensor changes to , and finally restore the flattened track tensor to its original shape , is the number of tracks in sensor A, is the number of tracks in sensor B; Step 3.2: Construct the local feature extraction layer. The local feature extraction layer first consists of a The convolution layer extracts the neighborhood information in the track tensor, fuses the neighbor track features, and sets its step size and padding to 1 to ensure that the matrix size does not change. Then, two The convolutional layer will The number of channels of the convolutional layer output feature is magnified by 4 times and then scaled back to the original size, and the track information of different channels is integrated. The change in the shape of the tensor in the convolutional layer is ; Step 3.3: Construct the global feature extraction layer. The global feature extraction layer consists of two global self-attention layers. First, the track tensor is flattened and its shape is given by Changes to , treat the flattened track tensor as a whole, and then calculate the query matrix, key matrix, and value matrix; Step 3.4: Construct row association mapping layer, column association mapping layer and global association mapping layer. The row association mapping layer is Convolutional layer, column-associative mapping layer is Convolutional layer, the global correlation map layer is a global self-attention layer that flattens the track tensor as input and restores the output tensor to its original shape with the same input and output dimensions; Step 3.5: Combine the outputs of the row association mapping layer, column association mapping layer, and global association mapping layer according to the channel dimension and pass a convolution kernel to , the convolution layer with a step size of 1, a filling of 0, and an output channel number of 1 is combined to perform association mapping to obtain the association matrix ; Step 4: Train and test the similarity segmentation network. The input of the network is the track tensor. , the output is the incidence matrix , according to the output correlation matrix The distribution of the median values ​​is used to obtain the associated results.

2. A track association method based on track similarity segmentation according to claim 1, characterized in that: The step 1 specifically includes the following sub-steps: Step 1.1: First, traverse all track points in each sensor and find the maximum and minimum values ​​of each dimension of the track coordinates in the current scene. Then, process each dimension separately. For each track point, subtract the corresponding minimum value from each dimension and divide it by the corresponding maximum minus minimum value to obtain a standardized, dimensionless normalized track. Step 1.2, traverse sensor A and get the number of tracks in sensor A , traverse sensor B and get the number of tracks in sensor B , intercept the nearest sampling points, the constructed shape size is The initial track tensor ,in All elements of are 0, and the constructed shape is The initial incidence matrix ,in All elements of are 0; Step 1.3: Arrange the tracks in sensor A and sensor B according to the track batch number from small to large, traverse sensor A, and obtain the track batch number in sensor A. Tracks , The shape is , traverse sensor B and obtain the first Tracks , The shape is ,Will and Flatten each row of , and splice the flattened tracks in the column direction to get the fused track , The shape is , in For splicing operation, and To flatten the track, the track will be fused later. Put in the initial track tensor No. Rank Column, get the track tensor ; Step 1.4, sort the tracks in sensor A and sensor B according to the track batch number from small to large, traverse sensor A, and obtain the first track in sensor A. Tracks , according to the associated annotations in the data set, we can get the Associated tracks ,in express At the position of the track of sensor B arranged from small to large according to the track batch number, the initial correlation matrix Middle Rank The column elements are set to 1 to obtain the true correlation matrix .

3. A track association method based on track similarity segmentation according to claim 1, characterized in that: The step 2 specifically includes the following sub-steps: Step 2.1: The total loss function is the weighted sum of similarity loss, row constraint loss, and column constraint loss. The total loss function is in is the weighting coefficient; Step 2.2: Similarity loss ensures that the generated correlation matrix is ​​as similar as possible to the true correlation matrix. In order to accurately measure the difference between elements in the matrix, the square error of each element in the correlation matrix obtained after similarity segmentation and the true correlation matrix is ​​used. The similarity loss is, in express No. Rank Column elements, express No. Rank Column elements, represents the 2-norm of the matrix; Step 2.3, the row constraint loss ensures that each row can have at most one associated pair, that is The sum of the elements in any row is less than or equal to 1 to prevent repeated associations and missed associations. The row constraint loss is, Step 2.4, the column constraint loss ensures that each column can have at most one associated pair, that is, The sum of the elements in any row is less than or equal to 1 to prevent repeated associations and missed associations. The column constraint loss is, 。 4. A track association method based on track similarity segmentation as claimed in claim 3, characterized in that: , , The specific value of is 1, 1, 1.

5. The track association method based on track similarity segmentation according to claim 1, characterized in that: In step 3.2, Convolutional layer and the first Adding layer normalization between convolutional layers speeds up network convergence. The convolutional layer and the second A GELU nonlinear activation function is added between convolutional layers.

6. A track association method based on track similarity segmentation according to claim 1, characterized in that: The step 3.3 specifically includes the following steps to calculate the query matrix, key matrix, and value matrix: in, 、 、 They are query matrix, key matrix, and value matrix respectively. 、 、 They are query mapping matrix, key mapping matrix, and value mapping matrix respectively. is the flattened track vector. In each global self-attention layer, 、 、 The shape is , 、 、 The shape is , The shape is , then calculate and The inner product of To make the training of the global self-attention layer more stable, we use SoftMax as the self-attention score. in is the self-attention score, with the shape of , and then calculate the self-attention score and The inner product of is output, in For output, a fully connected layer is used in each global attention layer to change its output dimension. The first layer enlarges the dimension to 4 times the original dimension, and the second layer scales it back to the original dimension. The GELU nonlinear activation function is connected between the two global self-attention layers.

7. The track association method based on track similarity segmentation according to claim 1, characterized in that: In step 3.4, the row association mapping layer uses Convolutional layer, whose stride is , filled with , the number of input and output channels is the same, and the column-associative mapping layer uses Convolutional layer, whose stride is , filled with , the number of input and output channels is the same.

Images