A method for associating interrupted tracks based on situation graph classification

Abstract

The present invention relates to an interrupted track association method based on situation diagram classification, and belongs to the field of radar target tracking technology. The present invention first performs an initial association on the new and old track pairs based on the time threshold and the speed threshold to reduce the total association time; then the track pairs that have passed the initial association are combined and converted into a situation diagram, and the position characteristics, speed characteristics and time characteristics of the track are pre-extracted; then a neural network model is constructed based on a residual network, and the track situation diagram is judged using the constructed network model; then the ambiguity problem of track association is solved based on 2D optimal allocation; finally, a simulation comparison experiment is used to verify that the proposed method has a significantly improved accuracy compared to several typical TSA algorithms currently available.

Description

Technical Field

[0001] The invention relates to an interrupted track association method based on situation graph classification, belonging to the technical field of radar target tracking. Background Art

[0002] Due to the reasons of target maneuvering, long radar sampling interval and low target detection probability, the detected target track is frequently interrupted, which seriously affects the subsequent command decision. The purpose of the interrupted track association method is to associate the new and old track segments belonging to the same target. The existing interrupted track association methods are mainly divided into traditional association methods and association methods combined with intelligent learning technology. The traditional association method has poor performance in dense interruption environment or long interruption time, and the association method combined with intelligent learning technology is not ideal in performance.

[0003] With the development of deep learning technology, deep learning methods based on image classification have shown powerful performance in target detection, classification and tracking. Therefore, a method for associating interrupted tracks based on situation graph classification is proposed. Combining radar data processing with deep learning technology, the reliability and stability of target tracking can be improved by accurately associating interrupted tracks, which has important application value. Summary of the invention

[0004] In order to solve the shortcomings of the above-mentioned prior art, a method for associating interrupted tracks based on situation diagram classification is provided. Firstly, the new and old track segments are initially associated based on the time threshold and the speed threshold to reduce the total association time; then the track pairs that have passed the initial association are combined and converted into a situation diagram, and the position characteristics, speed characteristics and time characteristics of the track are pre-extracted; then a neural network model is constructed based on the residual network, and the track situation diagram is judged using the constructed network model; then the ambiguity problem of track association is solved based on the two-dimensional optimal allocation; finally, the simulation comparison experiment is used to verify that the accuracy of the proposed method is significantly improved compared with several existing typical TSA association methods.

[0005] The invention discloses a method for associating interrupted tracks based on situation diagram classification, which is characterized in that it comprises the following steps:

[0006] Step 1: Divide the old track segment and the new track segment based on time, define the new track segment and the old track segment, construct the track association matrix, make an initial association judgment on the new and old track segments based on the time threshold and speed threshold, set the values ​​of the new and old track segments that pass the initial association in the track association matrix to 1, and set the rest to 0;

[0007] Step 2: Combine the old and new track segments that meet the initial association into one track segment in ascending order based on the time corresponding to each point track, record the corresponding track number, and mark the samples according to the prior association information, and combine all the new and old track segments that meet the initial association together; define three matrices, respectively with Location, The positions and velocities are used as rows, and the corresponding time information is used as columns to update the elements in the matrix, and the obtained matrix is ​​input into the TSASMC network model as a track situation diagram;

[0008] Step 3: Based on the residual network, a neural network model is constructed to make association judgments between the new and old track segments. The neural network model extracts the spatial feature information of the track situation map through multiple convolutional neural networks (CNNs), introduces residual connections to enhance the training efficiency and performance of the network, and uses a fully connected layer to summarize and classify these features to determine whether the new and old track segments belong to the same target.

[0009] Step 4: Input the track situation diagram into the trained neural network model. The output of the neural network model is a two-dimensional vector, which corresponds to the probability of not belonging to the same target and the probability of belonging to the same target. Update the value of the new and old track segments in the track association matrix corresponding to the combination of the initial association judgment; perform maximum value traversal on the matrix, select the non-zero and global maximum value, and judge the new and old track segments corresponding to the value as associated, and set the element value of the new and old track segments determined to be associated in the track association matrix to 0, and set the values ​​in the row and column corresponding to the value to 0, and continue to perform global maximization traversal until the maximum value obtained by the traversal is 0, and the final set is used as the final association result;

[0010] Step 5: Generate simulation data using the interrupted track environment construction method to compare and evaluate the correlation methods.

[0011] Preferably, the specific steps of step 1 are:

[0012] The old and new track segments are divided based on time. Assume that the maximum start time of all track segments obtained by statistics is , the minimum end time is , the new and old track segments are defined as follows:

[0013] Old track segment: The track segment whose end time is less than the maximum start time:

[0014] (1);

[0015] New track segment: A track segment whose start time is greater than the minimum end time:

[0016] (2);

[0017] In the formula for Time and track segment The state vector of and Represents the track segment The start and end time of and Respectively represent the number of old track segments and new track segments, ;

[0018] Based on the time threshold and speed threshold, the new and old track segments are initially associated with each other:

[0019] Time threshold: old track segment The end time of the new track segment is less than The starting time is as shown in formula 3:

[0020] (3);

[0021] Speed ​​threshold: old track segment Ending time With new track segment Start time The Euclidean distance between The ratio of the time difference is less than or equal to , as shown in Formula 4:

[0022] (4);

[0023] (5);

[0024] In the formula is the maximum value of the speed among all the track data obtained by statistics;

[0025] Based on the above initial correlation results, OK Track correlation matrix of columns , assign values ​​to the association matrix according to the initial association result, as shown in equations (6)-(7). After the new and old track segments that pass the initial association are first determined to be associated, their corresponding elements in the matrix are set to 1; the new and old track segments that do not pass the initial association are determined to be unassociated, and their corresponding elements in the matrix are set to 0. The track segments that pass the initial association are combined to be converted into a track situation diagram;

[0026] (6);

[0027] (7).

[0028] Preferably, the specific steps of step 2 are:

[0029] The old track segment that satisfies the initial association and new track segment The time points corresponding to each track are combined in ascending order to form a track segment. , record the corresponding track number , and mark the samples according to the prior association information, mark the ones belonging to the same target as 1, recorded as positive samples; mark the track segments that do not belong to the same target as 0, recorded as negative samples; when generating the validation set, all the new and old track segments that meet the initial association are combined together;

[0030] (8);

[0031] The combined track segments are normalized to adjust the track data to a uniform scale. of Location, The position and velocity information are globally normalized to The normalization of the position is shown in formula (9):

[0032] (9);

[0033] In the formula, ,

[0034] ;

[0035] For track segment Time information Perform local normalization, as shown in formula (10):

[0036] (10);

[0037] In the formula , .

[0038] Define three matrices , respectively Location, The position and speed are used as rows, and the time information is used as columns to update the elements in the matrix. The obtained matrix is ​​input into the TSASMC network model as a track situation diagram.

[0039] Set the matrix size to , initialize all elements of the matrix to 0, and update the factor , based on the update factor for the track segment of Location, The position, velocity and time are updated to obtain the corresponding coordinates in the matrix, as shown in formula (11):

[0040] (11);

[0041] In the formula Represents logarithmic value Round to the nearest integer. , Respectively represent the track after normalization The timing of the moment and the speed;

[0042] Based on the obtained coordinates, specific rows and columns of the matrix are updated, as shown in formula (12):

[0043] (12).

[0044] Preferably, the specific steps of step 3 are:

[0045] A neural network model is built based on the residual network to associate the new and old track segments. The neural network model extracts the spatial feature information of the track situation map through multiple convolutional neural networks (CNN), summarizes and classifies the features using a fully connected layer, and associates the new and old track segments that are determined to belong to the same target.

[0046] Use one The convolution layer performs preliminary feature extraction on the track situation map, as shown in formula (13):

[0047] (13);

[0048] in, Represents the track situation diagram, represents the kernel parameters of the initial convolutional layer, is the bias, is the output of the initial convolutional layer, Represents the convolution calculation, Normalize the data using a batch normalization layer, and The activation function performs nonlinear processing to obtain the output , as shown in formula (14):

[0049] (14);

[0050] in express Activation function, represents a batch normalization layer;

[0051] The neural network model includes 4 groups of residual blocks, each of which contains two basic residual units. There are two types of residual units: one does not contain Convolution is used to adjust the dimension, one is inclusive; each unit of the first group of residual blocks includes two The convolutional layer has 64 output channels, a stride of 1, and a padding of 1. Each convolutional layer is followed by a batch normalization layer and Activation function; the second group of residual blocks has 128 output channels and the first layer stride is 2. The convolution is dimensionally adjusted to achieve residual connection; the third and fourth groups set the number of output channels to 256 and 512 respectively, and also pass Convolution adjusts the dimension;

[0052] The outputs of the four residual blocks are shown in equations (15)-(20):

[0053] (15);

[0054] (16);

[0055] (17);

[0056] (18);

[0057] in:

[0058] (19);

[0059] (20);

[0060] in, represents the input of the residual unit, and They represent the first and second convolution kernel parameters of the residual unit that does not participate in dimension adjustment, and They represent the first and second convolution kernel parameters of the residual unit involved in dimension adjustment, Indicates the dimension adjustment Convolution kernel parameters, Indicates that As a residual block Input.

[0061] Finally, the model uses a global average pooling layer to average pool each feature map and output a vector of length 512, followed by a fully connected layer with an input dimension of 512 and an output dimension of 2. The output result is , as shown in formula (21):

[0062] (twenty one);

[0063] in, represents the weight of the fully connected layer, represents the global average pooling layer, represents the bias of the fully connected layer;

[0064] The model consists of 17 convolutional layers and 1 fully connected layer, a total of 18 weight layers, each convolutional layer is followed by a batch normalization layer and Activation function, during training and verification, the new and old track segments are combined to construct a track situation map, and the track situation map is input into the TSASMC network. Through training, the model outputs a two-dimensional vector , this vector represents the probability that the track segment combination belongs to different targets (labeled as 0) or the same target (labeled as 1), where represents the probability that the track segment combination belongs to different targets, It represents the probability that a combination of track segments belongs to the same target.

[0065] Preferably, the specific steps of step 4 are:

[0066] The track situation map is input into the trained neural network model. The output of the model is a two-dimensional vector, which corresponds to the probability of not belonging to the same target. The probability that and belong to the same target , the combination of the new and old track segments through the initial association is mapped to the track association matrix Update the value in:

[0067] (twenty two);

[0068] The set of track pairs for determining association is defined as , initialized to empty;

[0069] (twenty three);

[0070] Take the global maximum value of the track correlation matrix and set the old track segment corresponding to the maximum value With new track segment Track number combination Add to collection In , they are identified as track segments from the same target and associated:

[0071] (twenty four);

[0072] (25);

[0073] The associated new and old track segments are added to the track association matrix The element value in is set to 0, and the row and column values ​​corresponding to the value in the track association matrix are also set to 0. The global maximization traversal is continued until the maximum value obtained by the traversal is 0, and the final set is used as the final association result.

[0074] The invention provides an interrupted track association method based on situation diagram classification, and the beneficial effects are as follows:

[0075] 1. Convert the track data into a track situation diagram, so that the position characteristics, speed characteristics and time characteristics of the track can be more easily extracted.

[0076] 2. The interrupted track association problem is regarded as an image classification problem, and a neural network model is constructed based on the residual network to deal with the classification problem of the track situation map.

[0077] 3. An ambiguous judgment technology suitable for track situation diagram classification is designed.

[0078] Compared with several typical TSA algorithms currently available, the accuracy of the present invention is significantly improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] Figure 1 It is a track situation diagram belonging to the same target in the simulation scenario;

[0080] Figure 2 It is a track situation diagram of the objects that do not belong to the same target in the simulation scenario;

[0081] Figure 3 The dimension of the residual unit on the right is adjusted by adding a convolutional layer;

[0082] Figure 4 It is the model framework flow chart;

[0083] Figure 5 It is the correct association rate diagram of the association method under different numbers of tracks;

[0084] Figure 6 This is a graph showing the correct association rate of the association method under different interruption times. DETAILED DESCRIPTION

[0085] This embodiment describes in detail a method for associating interrupted tracks based on situation graph classification of the present invention, comprising the following steps:

[0086] Step 1: When the target track is interrupted, it will lead to a new batch. Each track segment corresponds to a batch number. TSA determines the new and old track segments belonging to the same target as associated, and the new and old track segments that do not belong to the same target as unassociated. In order to ensure that all new and old track segments belonging to the same target can be associated, the old track and the new track are divided based on time. Assume that the maximum start time corresponding to all track segments obtained by statistics is , the minimum end time is , the new and old track segments are defined as follows:

[0087] Old track segment: The track segment whose end time is less than the maximum start time:

[0088] (1);

[0089] New track segment: A track segment whose start time is greater than the minimum end time:

[0090] (2);

[0091] In the formula for Time Track The state vector of and Respectively indicate the track The start and end time of and Represent the number of old tracks and new tracks respectively, . It should be noted that the same track segment may belong to both the old track and the new track segment;

[0092] In order to reduce the association time and avoid the model from paying too much attention to irrelevant samples, most of the interrupted track segments that may be irrelevant are excluded, and the initial association judgment of the new and old track segments is performed based on the time threshold and speed threshold.

[0093] Time threshold: old track segment The end time of the new track segment is less than The starting time of . As shown in formula 3.

[0094] (3);

[0095] Speed ​​threshold: old track segment Ending time With new track segment Start time The Euclidean distance between The ratio of the time difference is less than or equal to . As shown in formula 4.

[0096] (4);

[0097] (5);

[0098] In the formula is the maximum speed of all track segment data obtained by statistics. OK Track correlation matrix of columns , assign values ​​to the association matrix according to the initial association result, as shown in Formula 6-7. After the new and old track segments that pass the initial association are first determined to be associated, their corresponding elements in the matrix are set to 1; the new and old track segments that do not pass the initial association are determined to be unassociated, and their corresponding elements in the matrix are set to 0. The track segments that pass the initial association are combined into a track situation diagram.

[0099] (6);

[0100] (7);

[0101] Step 2: ① Normalization: Normalize the old track segments that meet the initial association and new track segment The time points corresponding to each track are combined in ascending order to form a track segment. In order to facilitate the subsequent testing of the performance of the association method, the corresponding track number is recorded , and mark the samples according to the prior association information. Those belonging to the same target are marked as 1, recorded as positive samples; the combination of track segments that do not belong to the same target is marked as 0, recorded as negative samples. When generating the training set, in order to achieve a balanced number of sample categories, set an appropriate positive and negative sample ratio to combine the new and old track segments. When generating the validation set, in order to be closer to the real association scenario, all the new and old track segments that meet the initial association are combined together.

[0102] (8);

[0103] In the original data, the track segments are of different lengths and have a large position span, which is not conducive to the situation map conversion. In order to unify the data scale, the combined track segments are normalized. The normalization process adjusts the track segment data to a unified scale, thereby avoiding the adverse effects of too large or too small value range on the model training process. of Location, The position and velocity information are globally normalized to Take the position information as an example, as shown in Formula 9.

[0104] (9);

[0105] In the formula, ,

[0106] In this way, the track segment The position information and velocity information are globally normalized.

[0107] Considering the different lengths of each track segment, the large differences in the time range of the track segment and the need to maximize the use of time information of each track segment, while improving the resolution of the situation map, the track segment Time information Perform local normalization, as shown in formula 10:

[0108] (10);

[0109] In the formula , .

[0110] ② Situation diagram conversion: define three matrices , with the track segment time as the column, respectively Location, The elements in the matrix are updated with positions and velocities as rows, and the obtained matrix is ​​input into the TSASMC network model as a track situation diagram.

[0111] Set the matrix size to , initialize all elements of the matrix to 0, and update the factor , based on the update factor for the track segment of Location, The position, velocity and time are updated to obtain the corresponding coordinates in the matrix, as shown in formula 11:

[0112] (11);

[0113] In the formula Represents logarithmic value Round to the nearest integer. , Respectively represent the track after normalization The moment and speed of the moment.

[0114] Based on the obtained coordinates, specific rows and columns of the matrix are updated, as shown in Equation 12:

[0115] (12);

[0116] Step 3: A neural network model is constructed based on the residual network to associate the new and old track segments. The model extracts the spatial feature information of the track situation map through multiple convolutional neural networks (CNNs) and introduces residual connections to enhance the training efficiency and performance of the network. These features are summarized and classified using the fully connected layer, and the new and old track segments that are determined to belong to the same target are associated.

[0117] Use one The convolution layer (output channel number is 64, stride is 1, padding is 1) performs preliminary feature extraction on the track situation map.

[0118] (13);

[0119] in, Represents the track situation diagram, represents the kernel parameters of the initial convolutional layer, For bias. is the output of the initial convolutional layer, Represents the convolution calculation, Normalize the data using a batch normalization layer, and The activation function performs nonlinear processing to obtain the output .

[0120] (14);

[0121] in express Activation function, Represents a batch normalization layer.

[0122] The model includes 4 groups of residual blocks, each of which contains two basic residual units. There are two types of residual units: one does not contain Convolution performs dimensionality adjustment, one of which is included, such as Figure 3 As shown. Each unit of the first group of residual blocks consists of two The convolutional layer has 64 output channels, a stride of 1, and a padding of 1. Each convolutional layer is followed by a batch normalization layer and The second set of residual blocks has 128 output channels and a stride of 2. The convolution is dimensionally adjusted to achieve residual connection. The third and fourth groups set the number of output channels to 256 and 512 respectively, and also pass Convolution adjusts the dimension.

[0123] The outputs of the four residual blocks are shown in equations (15)-(18):

[0124] (15);

[0125] (16);

[0126] (17);

[0127] (18);

[0128] in:

[0129] (19);

[0130] (20);

[0131] in, represents the input of the residual unit, and They respectively represent the first and second convolution kernel parameters of the residual unit that does not participate in dimension adjustment. and They represent the first and second convolution kernel parameters of the residual unit involved in dimension adjustment, Indicates the dimension adjustment Convolution kernel parameters, Indicates that As a residual block Input.

[0132] Finally, the model uses a global average pooling layer to average pool each feature map and output a vector of length 512. Then there is a fully connected layer with an input dimension of 512 and an output dimension of 2. The output result is .

[0133] (twenty one);

[0134] in, represents the weight of the fully connected layer, represents the global average pooling layer, Represents the bias of the fully connected layer. Determine whether they belong to the same target based on the obtained results.

[0135] The model consists of 17 convolutional layers and 1 fully connected layer, a total of 18 weight layers, each convolutional layer is followed by a batch normalization layer and Activation function. During training and verification, the new and old track segments are combined to construct a track situation map, and the track situation map is input into the TSASMC network. Through training, the model outputs a two-dimensional vector This vector represents the probability that a combination of track segments belongs to different targets (labeled as 0) or the same target (labeled as 1), where represents the probability that the track segment combination belongs to different targets, represents the probability that the track segment combination belongs to the same target. The process framework of the model is as follows Figure 4 shown.

[0136] Step 4: Input the track situation map into the trained neural network model. The output of the model is a two-dimensional vector, which corresponds to the probability of not belonging to the same target. The probability that and belong to the same target . The new and old track segments that are initially associated are combined in the track association matrix Update the value in:

[0137] (twenty two);

[0138] The set of track pairs for determining association is defined as , initialized to empty.

[0139] (twenty three);

[0140] Take the global maximum value of the track correlation matrix and set the old track segment corresponding to the maximum value With new track segment Track number combination Add to collection In the process, they are identified as track segments coming from the same target and associated.

[0141] (twenty four);

[0142] (25);

[0143] The associated new and old track segments are added to the track association matrix Set the element value in to 0, continue the global maximization traversal until the maximum value obtained by the traversal is 0, and then stop. As the final result of the association.

[0144] Step 5: Generate simulation data using the interrupted track environment construction method mentioned in the literature (D. Ding, G. Wei, Z. Cao, et al. Simulation Environment Construction of Track Segment Association and Algorithm performance evaluation. Journal of System Simulation, 1-13 [2024-06-29].). When generating simulation data, consider the movement of the target in a two-dimensional plane, select uniform linear motion, uniform accelerated linear motion, uniform turning motion, and accelerated turning motion as the target motion model set, and the sensor samples the target 50 times with a sampling interval of 1s. Set the number of fixed state transitions of the target during this period to 5. Set the minimum speed of the target movement to , the maximum speed is , the maximum acceleration is , the maximum turning rate is The process noise in the motion process follows a Gaussian distribution with a mean of 0 and a standard deviation of 0.1. When the TSA simulation environment is generated, the initial position of the target is set to , the density control center coordinates are , the density control radius is .

[0145] Without considering the stationary state of the target relative to the radar, the interactive multiple model state-dependent mode transition probabilities Track Segment Association (IMMSDP-TSA) association method in the literature (S. Zhang, Y. Bar-Shalom. Tracksegment association for GMTI tracks of evasive move-stop-move maneuvering targets [J]. IEEE Transactions on Aerospace and Electronic Systems, 2011, 47(3): 1899-1914.), the statistical binary thresholds Track Segment Association (SBT-TSA) association method in the literature (L. Qi, H. Wang, Y. Liu. Track Segment Association Algorithm Based on Statistical Binary Thresholds [J]. Journal of Radars, 2015, 4(3): 301-308.), and the statistical binary thresholds Track Segment Association (SBT-TSA) association method in the literature (W. Xiong, P. Xu, Y. Cui, et al. Track segment association with dual contrast neural network[J]. IEEE Transactions on Aerospace and Electronic Systems, 2021, 58(1): 247-261.) and other three association methods were used as comparison objects. The correct association rate was selected as the performance indicator.

[0146] The interruption time is set to 10s. The effect of different track numbers on the performance of different TSA association methods is as follows: Figure 5As shown. The performance of the association method in this embodiment is better than the other three association methods. This embodiment has always maintained a high correct association rate. As the number of tracks increases, the association method in this embodiment decreases slightly but is still higher than the other three association methods. The correct association rates of the IMMSDP-TSA association method and the SBT-TSA association method remain at around 70%, and the performance of the TSADCNN association method is slightly lower. See Table 1 for specific data.

[0147] Table 1 Correct association rate of association methods under different numbers of tracks:

[0148] ;

[0149] The number of tracks is set to 50. The effects of different interruption times on the performance of different TSA association methods are shown in the figure. Figure 6 The association method of this embodiment has maintained good performance. When the interruption time is 30s, the performance decreases significantly. As the interruption time increases, the performance of the TSADCNN association method decreases significantly. The correct association rates of the IMMSDP-TSA association method and the SBT-TSA association method remain at around 70%. See Table 2 for specific data.

[0150] Table 2 Correct association rate of association methods under different interruption times:

[0151] ;

[0152] The above generally describes the present invention in detail, but it is obvious to a person skilled in the art that some modifications or improvements can be made to the present invention. Therefore, modifications or improvements that do not depart from the spirit of the present invention are within the scope of protection of the present invention.

Claims

1. A method for associating interrupted tracks based on situation graph classification, characterized in that The following steps are involved: Step 1: Divide the old track segment and the new track segment based on time, define the new track segment and the old track segment, construct the track association matrix, make an initial association judgment on the new and old track segments based on the time threshold and speed threshold, set the values ​​of the new and old track segments that pass the initial association in the track association matrix to 1, and set the rest to 0; Step 2: Combine the old track segments and new track segments that meet the initial association into one track segment in ascending order based on the time corresponding to each point track, record the corresponding track number, and mark the samples according to the prior association information, and combine all the new and old track segments that meet the initial association together; define three matrices, with x position, y position and speed as rows, and corresponding time information as columns to update the elements in the matrix, and input the obtained matrix into the TSASMC network model as a track situation diagram; Step 3: Based on the residual network, a neural network model is constructed to make association judgments between the new and old track segments. The neural network model extracts the spatial feature information of the track situation map through multiple convolutional neural networks (CNNs), introduces residual connections to enhance the training efficiency and performance of the network, and uses a fully connected layer to summarize and classify these features to determine whether the new and old track segments belong to the same target. Step 4: Input the track situation diagram into the trained neural network model. The output of the neural network model is a two-dimensional vector, which corresponds to the probability of not belonging to the same target and the probability of belonging to the same target. Update the value of the new and old track segments in the track association matrix corresponding to the combination of the initial association judgment; perform maximum value traversal on the matrix, select the non-zero and global maximum value, and judge the new and old track segments corresponding to the value as associated, and set the element value of the new and old track segments determined to be associated in the track association matrix to 0, and set the values ​​in the row and column corresponding to the value to 0, and continue to perform global maximization traversal until the maximum value obtained by the traversal is 0, and the final set is used as the final association result; Step 5: Generate simulation data using the interrupted track environment construction method to compare and evaluate the correlation methods; The specific steps of step 1 are: The old and new track segments are divided based on time. Assume that the maximum start time of all track segments obtained by statistics is The minimum end time is The new and old track segments are defined as follows: Old track segment: The track segment whose end time is less than the maximum start time: New track segment: A track segment whose start time is greater than the minimum end time: Where X i (k) is the state vector of track segment i at time k, and They represent the start and end time of the track segment i, I and J represent the number of old and new track segments, respectively. X i (k)=[x i (k),y i (k),v i (k)] T ; Based on the time threshold and speed threshold, the new and old track segments are initially associated with each other: Time threshold: The end time of the old track segment i is less than the start time of the new track segment j, as shown in formula (3): Speed ​​threshold: the end time k of the old track segment i e i and the start time k of the new track segment j s j The ratio of the Euclidean distance d between the positions and the time difference is less than or equal to V max , as shown in formula (4): Where V max is the maximum value of the speed among all the track data obtained by statistics; Based on the above initial association results, a track association matrix A with I rows and J columns is constructed. According to the initial association results, the association matrix is ​​assigned values, as shown in equations (6)-(7). After the new and old track segments that pass the initial association are first determined to be associated, their corresponding elements in the matrix are set to 1; the new and old track segments that do not pass the initial association are determined to be unassociated, and their corresponding elements in the matrix are set to 0. The track segments that pass the initial association are combined to be converted into a track situation diagram; The specific steps of step 2 are: The old track segment T that satisfies the initial association i and new track segment T j The time corresponding to each point track is combined in ascending order to form a track segment T ij , record the corresponding track number (i, j), and mark the samples according to the prior association information. The track segments belonging to the same target are marked as 1, which are recorded as positive samples; the track segments that do not belong to the same target are marked as 0, which are recorded as negative samples; when generating the validation set, all the new and old track segments that meet the initial association are combined together; The combined track segments are normalized to adjust the track data to a uniform scale. ij The x position, y position and velocity information are globally normalized, and the normalization of the x position is shown in formula (9): In the formula, For track segment T ij The time information t(k) is locally normalized, as shown in formula (10): In the formula Define three matrices G1, G2, and G3, with x position, y position, and velocity as rows, and time information as columns to update the elements in the matrices. The obtained matrices are input into the TSASMC network model as track situation diagrams. Set the matrix size to g×g, initialize all elements of the matrix to 0, update factor n = 1 / (g-1), and update the track segment T based on the update factor. ij The x position, y position, speed and time are updated to obtain the corresponding coordinates in the matrix, as shown in formula (11): Where round(a) means rounding the value a to the nearest integer, t(k) and v(k) respectively represent the time and speed of the track at time k after normalization; Based on the obtained coordinates, specific rows and columns of the matrix are updated, as shown in formula (12): The specific steps of step 3 are: A neural network model is built based on the residual network to associate the new and old track segments. The neural network model extracts the spatial feature information of the track situation map through multiple convolutional neural networks (CNN), summarizes and classifies the features using a fully connected layer, and associates the new and old track segments that are determined to belong to the same target. A 3×3 convolutional layer is used to perform preliminary feature extraction on the track situation map, as shown in formula (13): Among them, X represents the track situation diagram, represents the kernel parameter of the initial convolutional layer, b is the bias, is the output of the initial convolutional layer, Represents the convolution calculation, The data is normalized by the batch normalization layer and the ReLU activation function is used for nonlinear processing to obtain the output As shown in formula (14): Where rl represents the ReLU activation function and bn represents the batch normalization layer; The neural network model includes 4 groups of residual blocks, each of which contains two basic residual units. There are two types of residual units, one does not contain 1×1 convolution for dimension adjustment, and the other contains it. Each unit of the first group of residual blocks includes two 3×3 convolution layers, with 64 output channels, a stride of 1, and a padding of 1. Each convolution layer is followed by a batch normalization layer and a ReLU activation function. The second group of residual blocks has 128 output channels, a stride of 2 in the first layer, and 1×1 convolution is used to adjust the dimension to achieve residual connection. The third and fourth groups set the number of output channels to 256 and 512 respectively, and also adjust the dimension through 1×1 convolution. The outputs of the four residual blocks are shown in equations (15)-(18): in: Among them, X represents the input of the residual unit, and They represent the first and second convolution kernel parameters of the residual unit that does not participate in dimension adjustment, and They represent the first and second convolution kernel parameters of the residual unit involved in dimension adjustment, Represents the 1×1 convolution kernel parameters for dimension adjustment, and residualCell(x) represents taking x as the input of the residual block residualCell; Finally, the model uses a global average pooling layer to perform average pooling on each feature map and output a vector of length 512, followed by a fully connected layer with an input dimension of 512 and an output dimension of 2. The output result is z, as shown in formula (21): Among them, W f represents the weight of the fully connected layer, Avg represents the global average pooling layer, and b f represents the bias of the fully connected layer; The model includes 17 convolutional layers and 1 fully connected layer, a total of 18 weight layers. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function. During training and verification, the new and old track segments are combined to construct a track situation map, and the track situation map is input into the TSASMC network. Through training, the model outputs a two-dimensional vector [p0, p1], which represents the probability that the track segment combination belongs to different targets with a label of 0, or the same target with a label of 1, where p0 represents the probability that the track segment combination belongs to different targets, and p1 represents the probability that the track segment combination belongs to the same target.

2. A method for associating interrupted tracks based on situation diagram classification according to claim 1, characterized in that: The specific steps of step 4 are: The track situation diagram is input into the trained neural network model. The output of the model is a two-dimensional vector, which corresponds to the probability p0 of not belonging to the same target and the probability p1 of belonging to the same target. The value in the track association matrix A corresponding to the combination of the new and old track segments through the initial association is updated: Define the set of track pairs to be associated as Τ, which is initialized to be empty; Take the global maximum value of the track association matrix, add the track number combination (i, j) of the old track segment i and the new track segment j corresponding to the maximum value to the set Τ, and identify them as track segments from the same target for association: Τ=Τ+(i,j) (25) The element values ​​of the associated new and old track segments in the track association matrix A are set to 0, and the values ​​of the rows and columns corresponding to the values ​​in the track association matrix are also set to 0. The global maximization traversal is continued until the maximum value obtained by the traversal is 0, and the final set T is used as the final association result.

Images