Interrupted track association method and system based on track prediction, terminal and medium

Abstract

The invention belongs to the field of radar target tracking, and particularly relates to an interrupted track association method and system based on track prediction, a terminal and a medium, and the method comprises the steps: carrying out the coarse association of an old track and a new track, and carrying out the assignment of an association matrix according to a coarse association result; performing normalization processing on the track data of all the track pairs after coarse correlation; based on the preprocessed old flight path data, obtaining flight path data for predicting a new flight path through a pre-trained flight path prediction model based on a time sequence; on the basis of the preprocessed old track data and new track data, predicting the new track data, and obtaining an association classification result of the track pair through a pre-trained time sequence classification model; and according to an association classification result, assigning a value to the association matrix again, and obtaining an interrupted track association result based on the reassigned association matrix. According to the method, through time sequence classification and time sequence prediction, the interrupt track association accuracy under a relatively long interrupt duration is improved.

Description

Technical Field

[0001] The present invention belongs to the field of radar target tracking, and particularly relates to a method, system, terminal and medium for interrupted track association based on track prediction. Background Technique

[0002] Due to reasons such as platform movement, target maneuver, and long radar sampling intervals, track interruption phenomena often occur, which in turn cause serious interference to subsequent target tracking, situation awareness, information fusion, etc. Therefore, interrupted track association has become one of the key problems that urgently need to be solved in the field of radar data processing. The traditional TSA (Time Series Analysis) method first needs to obtain prior knowledge, assume a target motion model, and perform complex calculations on the interrupted track data, compare the differences between each track, and finally complete the association task through association assignment. However, when performing track prediction, there are often unreasonable assumptions, especially in long-term interruption and dense association environments, and the association effect is often unsatisfactory.

[0003] Currently, related technologies apply deep learning methods to the problem of interrupted track association. For example, a neural network model including a track feature extraction module, a time comparison module, a track feature comparison module, and a classifier module is used to calculate the association probability of two tracks. However, this method is applicable to shorter interruption durations and can achieve better association effects in the case of shorter interruption durations. The association effect for interrupted tracks with longer interruption durations is not ideal. Summary of the Invention

[0004] To solve the above problems, the present invention provides a method, system, terminal and medium for interrupted track association based on track prediction. A track prediction model based on time series is used for track prediction, and a time series classification model is used for association classification. Interrupted track association is performed through the combination of time series classification and time series prediction, improving the accuracy of interrupted track association in the case of longer interruption durations.

[0005] In the first aspect, the technical solution of the present invention provides a method for interrupted track association based on track prediction, including the following steps: Perform rough association on all old tracks and new tracks in the scene based on time matching and speed matching, and assign values to the association matrix according to the rough association results; Normalize the track data of all track pairs after rough association; Preprocess the track data of old tracks and new tracks. The preprocessing includes mean filtering and variable dimension expansion operations; Based on the preprocessed old track data, obtain the track data of the predicted new track through a pre-trained track prediction model based on time series; Based on the preprocessed old track data, new track data, and predicted new track data, obtain the association classification result of the track pair through a pre-trained time series classification model; Reassign the association matrix according to the association classification result, and obtain the interrupted track association result based on the reassigned association matrix.

[0006] In an alternative embodiment, based on the preprocessed old track data, obtain the track data of the predicted new track through a pre-trained track prediction model based on time series, specifically including: Perform dimensional transformation on the input data through the first linear layer; the input data refers to the preprocessed old track data; The output data of the first linear layer flows into the Mamba model and the reverse Mamba model respectively, and processes the sequence data from the forward and reverse directions respectively; Merge the output data of the Mamba model and the reverse Mamba model through an addition operation; Perform normalization processing on the merged data through the first normalization layer; Perform non-linear feature extraction on the output data of the first normalization layer through the feed-forward network layer; Perform an addition operation on the output data of the first normalization layer and the feed-forward network layer and then input it into the second normalization layer for normalization processing; Input the output data of the second normalization layer into the second linear layer, and output the track data of the predicted new track by the second linear layer.

[0007] In an alternative embodiment, the objective function of the track prediction model in the training phase is to use the contrastive loss to calculate the loss value between the track data of the predicted new track and the actual new track, defined as:

[0008] Among them, is a binary label. If the two track segments of the old track and the actual new track in the sample come from the same target, then , otherwise ; represents the interval distance, which is the shortest distance between two track segments belonging to different targets in the sample; represents the Euclidean distance between the predicted new track and the actual new track.

[0009] In an alternative embodiment, based on the preprocessed old track data, new track data, and predicted new track data, obtain the association classification result of the track pair through a pre-trained time series classification model, specifically including: Perform a subtraction operation on the preprocessed old track data and new track data; Perform a splicing operation on the subtraction operation result, the new track data, and the predicted new track data; Input the splicing result into a pre-trained time series classification model, and the time series classification model outputs the association classification result of the track pair.

[0010] In an alternative embodiment, the objective function of the time series classification model in the training phase is to use the cross-entropy loss function to calculate the loss value between the predicted probability and the true distribution. The output of the time series classification model for each sample is , is the probability that the old track and the new track do not belong to the same target, is the probability that the old track and the new track belong to the same target. The true label is 0 or 1, and the corresponding cross-entropy loss function is expressed as:

[0011] where is the number of samples; The total loss function is expressed as: .

[0012] In an alternative embodiment, input the splicing result into a pre-trained time series classification model, and the time series classification model outputs the association classification result of the track pair, specifically including: Let the splicing result flow into the increasing filter, the decreasing filter, the peak filter, and the first Inception module respectively; Perform a splicing operation on the output data of the increasing filter, the decreasing filter, the peak filter, and the first Inception module to obtain the integrated feature; The integrated feature is passed through the second Inception module twice to extract multi-scale features; Perform the first residual connection on the input data and the extracted multi-scale features; The result of the first residual connection is passed through the third Inception module three times to extract multi-scale features again; Perform the second residual connection on the result of the second residual connection and the multi-scale features extracted again; Reduce the feature dimension of the result of the second residual connection through the global average pooling layer; Pass the output data of the global average pooling layer through the linear layer to output the association classification result of the track pair.

[0013] In an alternative embodiment, reassign the association matrix according to the association classification result, and obtain the interrupted track association result based on the reassigned association matrix, specifically including: The association classification result includes the th old track and the A new track The probability of not belonging to the same target and the probability of belonging to the same target ; The element in the th row and th column of the association matrix is the th old track and the th new track The association mark between them ; Indicates that the two are not associated, Indicates that the two are associated; When and the current association mark is , update the association mark to 0; Define the set and initialize it to be empty, which is used to store the serial numbers of the newly and old tracks that are finally associated; Traverse the track pairs in the association matrix that satisfy , find the serial number of the old track with the largest value and the new track in each group of track pairs, and mark it as ; Add the mark to the set , and set all the elements in the th row and th column of the association matrix to 0, and then traverse the track pairs in the association matrix that satisfy again; finally, the content in the set is the interrupted track association result.

[0014] In a second aspect, the technical solution of the present invention provides an interrupted track association system based on track prediction, including: A track rough association unit, which is used to perform rough association on all old tracks and new tracks in the scene based on time matching and speed matching, and assign values to the association matrix according to the rough association result; A track data normalization processing unit, which is used to perform normalization processing on the track data of all track pairs after rough association; A track data preprocessing unit, which is used to perform preprocessing on the track data of old tracks and new tracks, and the preprocessing includes mean filtering and variable dimension expansion operations; A track data prediction unit, which is used to obtain the track data of the predicted new track based on the preprocessed old track data through a pre-trained track prediction model based on time series; An associated classification unit, configured to obtain an associated classification result of a track pair through a pre-trained time series classification model based on pre-processed old track data, new track data, and predicted new track data; An associated result acquisition unit, configured to re-assign values to an association matrix according to the associated classification result to obtain an interrupted track association result.

[0015] In a third aspect, a technical solution of the present invention provides a terminal, including: A memory, configured to store an interrupted track association program based on track prediction; A processor, configured to implement the steps of the interrupted track association method based on track prediction as described in any one of the above when executing the interrupted track association program based on track prediction.

[0016] In a fourth aspect, a technical solution of the present invention provides a computer-readable storage medium, on which an interrupted track association program based on track prediction is stored. When the interrupted track association program based on track prediction is executed by a processor, the steps of the interrupted track association method based on track prediction as described in any one of the above are implemented.

[0017] As can be seen from the above technical solutions, the present application has the following advantages: First, the new and old tracks are roughly associated, and the track data of the track segments are normalized and pre-processed. Then, through a track prediction model based on time series, the old track is used to predict the new track. Next, according to the prediction result and the track data of the actual track segment, an association classification is performed through a time series classification model. Finally, an interrupted track association is obtained based on the classification result. The present invention uses a track prediction model based on time series to perform track prediction, which can fully learn the time series characteristics and motion laws in the old track. The prediction method based on historical data not only considers the motion inertia of the target but also can capture the trend changes of the target's motion. Through in-depth analysis of the old track, the prediction model can more accurately predict the possible direction and state of the new track. And a time series classification model is used for association classification, which can comprehensively analyze factors such as the similarity, difference, and matching degree of time series characteristics between the predicted track and the actual track, and then accurately judge whether there is an association relationship between track pairs, avoiding association errors caused by unreasonable simple assumptions of the target motion model, and being able to more flexibly adapt to different target motion patterns and complex environmental conditions, thus significantly improving the accuracy of track association judgment. In the case of a long interruption duration, the time series prediction model uses the historical information of the old track to accurately infer the motion track of the target during the interruption; the time series classification model can analyze the association relationship between tracks based on the prediction result and the actual track data. The two cooperate with each other, so that in the scenario of a long interruption duration, a high track association accuracy can still be maintained, and the accuracy of interrupted track association in the case of a long interruption duration is improved. Brief Description of the Drawings

[0018] To more clearly illustrate the technical solutions of the present application, the accompanying drawings required for the description will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, other accompanying drawings can be obtained based on these drawings without creative efforts.

[0019] Figure 1 It is a schematic flowchart of a method for associating interrupted tracks based on track prediction provided by an embodiment of the present invention.

[0020] Figure 2 It is a schematic diagram of the framework structure for associating interrupted tracks based on track prediction.

[0021] Figure 3 It is a schematic diagram of the change of track data.

[0022] Figure 4 It is a schematic diagram of the S-mamba model architecture.

[0023] Figure 5 It is a schematic diagram of the Mamba model architecture.

[0024] Figure 6 It is a schematic diagram of the H-Inception Time model architecture.

[0025] Figure 7 For Figure 6 the schematic diagram of the filter structure created manually in

[0026] Figure 8 For Figure 6 the schematic diagram of the Inception module structure in

[0027] Figure 9 It is a schematic diagram of the complete track of 5000 targets including noise.

[0028] Figure 10 It is a schematic block diagram of the structure of a system for associating interrupted tracks based on track prediction provided by an embodiment of the present invention.

[0029] Figure 11 It is a schematic diagram of the structure of a terminal provided by an embodiment of the present invention. Detailed Embodiments

[0030] To make the application purpose, features, and advantages of this application more obvious and understandable, the following will use specific embodiments and accompanying drawings to clearly and completely describe the technical solutions protected by this application. Obviously, the embodiments described below are only a part of the embodiments of this application, rather than all embodiments. Based on the embodiments in this patent, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of this patent.

[0031] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which this invention belongs. The terms used in the specification of this invention herein are only for the purpose of describing specific embodiments and are not intended to limit this invention.

[0032] Inception module: Consists of multiple parallel convolutional layers of different sizes, capable of capturing information at different spatial scales, enabling the model to learn rich representations of the input data and effectively capturing local and global patterns in the data.

[0033] Mamba model: A new type of State Space Model (SSM), mainly developed by Albert Gu (Carnegie Mellon University) and Tri Dao (Princeton University).

[0034] Figure 1 This is a schematic flowchart of a method for associating interrupted tracks based on track prediction provided by an embodiment of the present invention. Among them, Figure 1 The execution subject can be a system for associating interrupted tracks based on track prediction. The method for associating interrupted tracks based on track prediction provided by an embodiment of the present invention is executed by a computer device. Correspondingly, the system for associating interrupted tracks based on track prediction runs in the computer device. According to different requirements, the order of steps in this flowchart can be changed, and some can be omitted.

[0035] As Figure 1 shown, this method includes the following steps.

[0036] S1, perform rough association on all old tracks and new tracks in the scene based on time matching and speed matching, and assign values to the association matrix according to the rough association results.

[0037] S2, perform normalization processing on the track data of all track pairs after rough association.

[0038] S3, perform preprocessing on the track data of old tracks and new tracks. The preprocessing includes mean filtering and variable dimension expansion operations.

[0039] S4. Based on the preprocessed old track data, obtain the track data of the predicted new track through a pre-trained track prediction model based on time series.

[0040] S5. Based on the preprocessed old track data, new track data, and predicted new track data, obtain the correlation classification result of the track pair through a pre-trained time series classification model.

[0041] S6. Reassign the correlation matrix according to the correlation classification result, and obtain the interrupted track correlation result based on the reassigned correlation matrix.

[0042] In this embodiment, a track prediction model based on time series is used for track prediction, which can fully learn the time series features and motion laws in the old track. The prediction method based on historical data not only considers the motion inertia of the target but also can capture the trend changes of the target's motion. Through in-depth analysis of the old track, the prediction model can relatively accurately predict the possible direction and state of the new track. And a time series classification model is used for correlation classification, which can comprehensively analyze factors such as the similarity, difference, and matching degree of time series features between the predicted track and the actual track, and then accurately judge whether there is a correlation relationship between track pairs, avoiding correlation errors caused by unreasonable simple assumption of the target motion model, and being able to more flexibly adapt to different target motion patterns and complex environmental conditions, thus significantly improving the accuracy of track correlation judgment. In the case of a long interruption duration, the time series prediction model uses the historical information of the old track to accurately infer the motion track of the target during the interruption; the time series classification model can analyze the correlation relationship between tracks based on the prediction result and the actual track data. The two cooperate with each other, so that in the scenario of a long interruption duration, a high track correlation accuracy can still be maintained, and the interrupted track correlation accuracy in the case of a long interruption duration is improved.

[0043] Further, as a refinement and extension of the specific implementation manner of the above embodiment, in order to fully illustrate the specific implementation process in this embodiment, another method for associating interrupted tracks based on track prediction is provided. This method first filters out new and old tracks that are less likely to be associated through rough association, performs mean filtering on the remaining new and old tracks to make the track data "smooth", and at the same time performs first-order difference for variable dimension expansion. The result after the difference is spliced with the previous data to increase the variable dimension. Then, based on the Mamba model, the old track is used to predict the new track, and the prediction result is compared with the new track, so that the distance between the new and old tracks determined to be associated becomes smaller, and the distance between the new and old tracks determined not to be associated becomes larger. Finally, the difference between the prediction result and the new track and the new and old tracks before prediction is spliced, and the spliced result is classified through the H-Inception model to obtain the probability of whether the new and old tracks are associated. This method includes the following steps.

[0044] SS1 coarsely correlates all the old tracks and new tracks in the scene based on time matching and speed matching, and assigns values to the association matrix according to the coarse correlation results.

[0045] When a target's track is interrupted, multiple track segments may be generated for the same target. In a scene of interrupted track association, it is possible that multiple targets' tracks are interrupted. The task of interrupted track association is to pair and associate the temporally adjacent track segments belonging to the same target in this scene. In the original track data, each track contains information in multiple dimensions. Among them, the position information, position information, speed scalar information, and corresponding time information of the target are selected as the feature information for track association, that is, track data. The moment corresponding to the first point measurement sampled by a track segment is defined as the start moment of this track segment, and the moment corresponding to the last point measurement is defined as the end moment of this track segment. When the end moment of a track segment is before the start moment of another track segment, the former is defined as the old track, and the latter is defined as the new track. A pair of old track and new track that needs to be judged whether they are associated is called a track pair. The start and end moments of all track segments in the scene are statistically analyzed to obtain the maximum start moment and the minimum end moment of all track segments as . The definitions of new tracks and old tracks are as follows: Old track: A track segment whose end moment is less than .

[0046] (1) New track: A track segment whose start moment is greater than .

[0047] (2) where represents the th old track, represents the th new track, represents the state information of the th old track at moment, represents the th new track at moment, and respectively represent the start moment and end moment of the th old track, and respectively represent the start moment and end moment of the th new track, and represent the number of old and new tracks respectively.

[0048] Define row and column association matrix .

[0049] (3) In the formula when it means to judge the th old track and the th new track are associated. when it means to judge the th old track and the th new track are not associated. After calculating with the interrupted track association algorithm introduced in this embodiment, each old track is associated with at most one new track, and each new track is associated with at most one old track, that is, each row (each column) of the association matrix has at most one element as 1, and the rest are 0.

[0050] Only discuss the situation where the target is in the two-dimensional rectangular coordinate system. When the old track and the new track are the front and back two track segments interrupted on the same track of the same target, then and must satisfy the matching relationship in terms of time and speed, that is, the start time of the new track is greater than the end time of the old track , and the ratio of the interruption distance to the interruption time is less than or equal to the maximum movement speed of the target. In order to reduce the association time, avoid the model from over-focusing on irrelevant samples, and at the same time exclude most of the interrupted track segments that are unlikely to be associated, the initial association judgment is made on all old tracks and new tracks in the scene based on time matching and speed matching respectively. Based on the rough association result, the association matrix is assigned values. If the old track and the new track simultaneously satisfy time matching and speed matching, then let the old track and the new track correspond to the in the association matrix , otherwise let .

[0051] (4) Formula (4) represents the track pairs that simultaneously satisfy time matching and speed matching. and respectively represent the The position of the end moment of the old track in direction and direction, and respectively represent the position of the start moment of the th new track in direction and direction. Since it is difficult to obtain the prior information of the target, the information of all track segments in the scene is statistically analyzed, and the maximum value of the speed scalar obtained is used as .

[0052] SS2, and the track data of all track pairs after rough association are normalized.

[0053] To avoid numerical problems and enhance the stability of the model, the track data is normalized. During the normalization process, for position information, position information, speed scalar information and time information are normalized using the maximum and minimum values in the global range. The normalization standard formula is: (5) In the formula , , and respectively represent the of the track pair at the moment position information, position information, speed scalar information and time information; , , and respectively represent the of the track pair after normalization at the moment position information, position information, speed scalar information and time information. The information of each dimension of all track segments in the scene is statistically analyzed. and respectively represent the minimum and maximum values of the target obtained statistically in the direction; and respectively represent the minimum and maximum values of the target obtained statistically in the direction; and respectively represent the minimum and maximum values of the speed scalar of the target obtained statistically. and respectively represent the minimum and maximum values of the time information of the target obtained by statistics.

[0054] SS3, track prediction and association classification.

[0055] The traditional TSA algorithm, through different track prediction strategies, extends the old track to the new track, or predicts both the old track and the new track to the same time point to judge the similarity of the two tracks. In a non-dense environment, this method can achieve track association more accurately. However, in a complex environment, due to the assumptions used in the track prediction process being difficult to fully adapt to the actual situation, the association accuracy drops significantly. This embodiment draws on the idea of the traditional algorithm to measure the similarity of track pairs through track prediction, and constructs a track prediction-based interrupted track association framework based on deep learning theory to solve the association problem of traditional interrupted tracks. The framework measures the similarity between the old track and the new track by predicting the old track to the new track and making a comparison, and judges whether the old and new tracks belong to the same target based on this. As Figure 2 shown, the framework mainly consists of three modules: an initial module, a track prediction module, and a classification module.

[0056] SS3.1, use the initial module to preprocess the track data of the old track and the new track.

[0057] The main function of the initial module is to smooth the track data and achieve feature dimension elevation. In interrupted track association, it is not only necessary to measure the similarity of the features of the old and new tracks, but also to analyze the differences between the two. After the initial module, by subtracting the variables of the extended old and new track data one by one, the difference features between the two track segments can be obtained. Subsequently, this difference is concatenated with the prediction result of the old track and the new track in the variable dimension, comprehensively considering the similarity and difference between the track segments, and the concatenated result is input into the classification module to finally obtain the probability of whether the track pair is associated.

[0058] The input track data usually contains observation noise, so it appears relatively "rough" and is difficult to effectively extract features. Using such "rough" data for track prediction will significantly reduce the prediction accuracy and have a direct impact on the final association performance. The traditional interrupted track association algorithm usually uses the interactive multiple model algorithm based on Kalman filtering to track and filter the track data before track prediction to obtain a "smooth" track that is more in line with the true trajectory of the target. However, the processing flow of this method is relatively complex and highly dependent on the acquisition of prior information. In contrast, deep learning algorithms have strong feature extraction capabilities and can directly learn the internal pattern of target motion from track data without relying on prior information. Therefore, good results can be obtained as long as the track data is ensured to be sufficiently "smooth". This embodiment uses mean filtering to "smooth" the input track data, and its calculation formula is: (6) Among them, is the mean filter window size. For track data with a long time step, a larger filter window can be selected.

[0059] Given that the characteristic dimension of the input track data is small, in this embodiment, the first-order difference method is used to obtain new features, and its calculation formula is: (7) Subsequently, the first row of the filtered track data is removed, and it is concatenated with the difference result in the feature dimension. The track data processed through the above operations has a reduced time step length compared to the input track data by , but its feature dimension doubles, thereby enhancing the feature extraction ability of the model. The shape change of the track data is as shown in Figure 3 . Figure 3 In represents the time step length of the original track data, represents the number of features of the original track data.

[0060] SS3.2. Based on the preprocessed old track data, the track data of the predicted new track is obtained through a pre-trained track prediction model based on time series.

[0061] When performing track prediction, the entire track is regarded as a multivariate time series, and the old track is predicted and trained with the entire new track as the target, so that the prediction module can accurately map the old track to the new track. For associated new and old track pairs, the Euclidean distance between the prediction result of the old track and the new track on each variable should be as small as possible to ensure that the prediction result is closer to the new track; for unassociated track pairs, the Euclidean distance between the prediction result of the old track and the new track is relatively large.

[0062] In this embodiment, the S-Mamba model is selected as the track prediction module in the framework, and its overall structure is as shown in Figure 4 . Among them, Mamba, as the core component of the S-Mamba model, is constructed based on the Selective State Space Model (Selective SSM), aiming to efficiently process sequence dependencies while maintaining an approximately linear time complexity. This model performs excellently in capturing relevant information, especially having significant advantages in processing ultra-long time series data. Its model structure is as shown in Figure 5 .

[0063] Among them, Selective SSM is based on the State Space Model (SSM). By parameterizing the input, the model can selectively process information. SSM is a mathematical method for time series modeling, mainly composed of a state equation and an output equation. Its specific form is as follows.

[0064] (8) (9) Among them, represents the latent state representation at a given time , represents the input at a given time , represents the output at a given time . A, B, C, and D represent the state transition matrix, input matrix, output matrix, and direct transmission matrix respectively.

[0065] Based on the track prediction model, the track data of the predicted new track is obtained through the following specific steps.

[0066] SS3.2.1, perform dimensional transformation on the input data through the first linear layer; the input data refers to the preprocessed old track data.

[0067] SS3.2.2, the output data of the first linear layer flows into the Mamba model and the reverse Mamba model respectively, and processes the sequence data from the forward and reverse directions respectively.

[0068] SS3.2.3, merge the output data of the Mamba model and the reverse Mamba model through an addition operation.

[0069] SS3.2.4, normalize the merged data through the first normalization layer.

[0070] SS3.2.5, perform non-linear feature extraction on the output data of the first normalization layer through a feed-forward network layer.

[0071] SS3.2.6, perform an addition operation on the output data of the first normalization layer and the feed-forward network layer, and then input it into the second normalization layer for normalization processing.

[0072] SS3.2.7, input the output data of the second normalization layer into the second linear layer, and the second linear layer outputs the track data of the predicted new track.

[0073] SS3.3. Based on the preprocessed old track data, new track data, and predicted new track data, obtain the association classification result of the track pair through a pre-trained time series classification model.

[0074] By using the trained track prediction module to predict the expanded old track data, obtain its prediction result relative to the corresponding new track. For a pair of new and old tracks that can be determined to be associated, the difference between their new and old track data should be significantly different from that of non-associated track pairs. At the same time, the similarity between the prediction result of the old track and the new track should be higher than that of non-associated track pairs.

[0075] In order to comprehensively express the difference and similarity between track pairs, in this embodiment, the difference between the expanded new and old track data, the prediction result, and the new track data are concatenated in the feature dimension to form a time series. The main function of the classification module is to use the time series classification model to extract features and classify this series. Through training, this module can classify and judge them by combining the difference and similarity between track pairs. For a pair of new and old tracks to be determined, the more similar the prediction result of the old track is to the new track, the higher the probability that this track pair is determined to be associated. Combining the difference features of the track pair, the classification module can more accurately judge the association relationship between the new and old tracks.

[0076] During the movement of the target, it usually maintains a fixed movement state for a period of time, making the corresponding track data show a specific change trend during this period. These changes may be manifested as increasing or decreasing. The InceptionTime model is a deep learning network that combines the Inception module and residual connections, specifically designed to process time series data. The H-InceptionTime model introduces three non-learnable one-dimensional general filters on the basis of the InceptionTime structure, which are used to detect the upward trend, downward trend, and peak value in the time series respectively, so as to effectively solve the time series classification problem. Based on this, this embodiment selects the H-InceptionTime model as the time series classification model in the classification module, and its overall model structure is as Figure 6 shown. Among them, the specific structure of the manually created filter part is as Figure 7 shown, and the specific structure of the Inception module is as Figure 8 shown.

[0077] Based on the time series classification model, the association classification of track pairs specifically includes the following steps.

[0078] SS3.3.1. Perform a subtraction operation on the preprocessed old track data and new track data.

[0079] SS3.3.3, perform a splicing operation on the subtraction operation result, the new track data, and the predicted new track data.

[0080] SS3.3.3, input the splicing result into a pre-trained time series classification model, and the time series classification model outputs the association classification result of the track pair.

[0081] SS3.3.3.1, make the splicing result flow into the increasing filter, the decreasing filter, the peak filter, and the first Inception module respectively.

[0082] SS3.3.3.2, perform a splicing operation on the output data of the increasing filter, the decreasing filter, the peak filter, and the first Inception module to obtain the integrated features.

[0083] SS3.3.3.3, the integrated features are extracted by the second Inception module twice to obtain multi-scale features.

[0084] SS3.3.3.4, perform the first residual connection on the input data and the extracted multi-scale features.

[0085] SS3.3.3.5, the result of the first residual connection is extracted by the third Inception module three times to obtain multi-scale features again.

[0086] SS3.3.3.6, perform the second residual connection on the result of the second residual connection and the multi-scale features extracted again.

[0087] SS3.3.3.7, reduce the feature dimension of the result of the second residual connection through the global average pooling layer.

[0088] SS3.3.3.8, the output data of the global average pooling layer passes through the linear layer to output the association classification result of the track pair.

[0089] In order to distinguish the track segments of different targets, it is necessary to design corresponding objective functions to optimize the network parameters. The higher the similarity between two track segments, the smaller the distance between them. The goal of the interrupted track association framework based on track prediction is: for two track segments determined to be associated, make the prediction result of the old track closer to the corresponding new track; for two track segments determined not to be associated, make the prediction result of the old track more different from the corresponding new track. In addition, it is also necessary to ensure that the model can accurately classify the spliced time series. Based on the above considerations, this embodiment designs two objective functions: one is for the track prediction model to optimize the distance between the prediction result of the old track and the new track; the other is for the time series classification to optimize the classification result of the model.

[0090] 1) Distance optimization: Use contrastive loss to calculate the loss value between the prediction result of the old track and the new track during the training phase. Its definition is as follows: (10) where is a binary label. If the two track segments of the old track and the actual new track in the sample come from the same target, then ; otherwise . represents the interval distance, which is the shortest distance between two track segments belonging to different targets in the sample. represents the Euclidean distance between the predicted new track and the actual new track.

[0091] 2) Classification result optimization: Use the cross-entropy loss function to calculate the loss value between the model prediction probability and the true distribution. For the case of containing samples, assume the output of the model for each sample is , and the true label is 0 or 1. Then the corresponding cross-entropy loss function can be expressed as: Its definition is as follows: (11) The total loss function is described as follows: (12) SS4, reassign the association matrix according to the association classification result, and obtain the interrupted track association result based on the reassigned association matrix.

[0092] Update the values in the track association matrix according to the output result of the time series classification model. For the track pair , the output of the model is the probability that the old track and the new track do not belong to the same target and the probability that they belong to the same target. When and the current association flag , the model determines that this track pair is unassociated and updates its value in the association matrix from 1 to 0, as shown in Equation (13) (13) Define the set as , which is used to store the sequence numbers of the newly and old tracks that are finally associated, and is initialized to be empty.

[0093] (14) After the time series classification model finishes judging, traverse the track pairs in the track association matrix that meet the condition , and find in each group of track pairs And the old track with the largest value and the new track . Mark it as , and add this mark to the set . At this time, regard the old track and the new track as consecutive track segments of the same target. After selecting each pair of associated tracks, update the values of the corresponding elements in the association matrix , and set all the element values in the th row and the th column of the association matrix to 0. As shown in Eqs. (15)-(17).

[0094] (15) (16) (17) Continue to perform a global maximum value traversal until the maximum value obtained from the traversal is less than or equal to 0 and stop. Finally, the content in the set is the result of track association.

[0095] The following explains network training and parameter selection.

[0096] During network training, selecting an appropriate interval distance helps to better perform distance optimization. Based on this, an interval distance analysis experiment is carried out. Use the average recall rate and the optimal recall rate to represent the average performance and the optimal performance of the algorithm during training respectively. Assume that N rounds of training are carried out. The definition of (18) where K is the number of correctly associated new and old track pairs, is the number of correctly associated new and old track pairs by the algorithm after the th training ends. The definition of (19) Data preparation: Generate a data set. Considering the movement of the target in a two-dimensional plane, select the CV, CA, CT, and CAT models as the set of simulation movement models. The Markov state transition matrix between the target movement models is as shown in Eq. 20.

[0097] (20) Set the process noise during the target movement process to follow a Gaussian distribution with a mean of 0 and a standard deviation of 0.1, and the sensor position coordinates are , angular measurement error , ranging error , the maximum speed of the target is , the minimum speed is , the maximum acceleration is , the minimum acceleration is , the maximum turning rate is , the minimum turning rate is , the total sampling duration is 100 s, the sampling interval is 1 s, and the initial position of the target is randomly generated , the initial heading is randomly generated , the initial velocity is randomly generated . By setting the density control radius , and the density control center coordinates are , a radar measurement area of size is obtained. Set the number of motion state switches of each target during this period to 5, and generate the complete tracks of 5000 targets, as shown in Figure 9 .

[0098] Interval distance analysis: The interval distance is a key hyperparameter affecting the performance of TSAF-TP, which determines the minimum distinguishable degree between different target track segments. During the model training process, adjusting the value of will directly affect the discriminant ability of the model for similar track segments, thus affecting the classification accuracy and generalization performance. If 𝑚 is set too small, the model may not be able to effectively distinguish the track segments of different targets; if set too large, the track segments that should be associated may be misjudged as irrelevant. Therefore, it is necessary to evaluate different interval distances (0.4, 0.6, 0.8, 1.0, 1.2) to determine the optimal value.

[0099] Select the complete tracks of 500 targets for verification analysis, and set the interruption duration to 10 s. The experimental results of the average recall rate and the optimal recall rate of the model under different interval distances are shown in Table 1.

[0100] Table 1: Association performance under different interval distances

[0101] As can be seen from Table 1, when the interval distance is 1.0, the performance of TSAF-TP is optimal. Then the interval distance of 1.0 will be used in subsequent experiments.

[0102] After that, algorithm verification and analysis are carried out. First, the network anti-noise performance is tested. It is very difficult to obtain an ideal noise-free dataset in a real environment. To explore the anti-noise performance of the method in this embodiment, the association effects in different noise scenarios are compared, and the noise levels in the real environment are simulated by setting different angle measurement errors and ranging errors.

[0103] Select the complete tracks of 500 targets for verification and analysis. The set interruption duration is 10 s. The experimental results of the average recall rate and the optimal recall rate of the model under different noise levels are shown in Table 2.

[0104] Table 2: Association performance under different noises

[0105] As can be seen from Table 2, the method in this embodiment shows good performance under different noises, thus proving that the method in this embodiment has good anti-noise performance.

[0106] The comparative experiment is as follows: To describe the performance of the algorithm, the correct track association rate is selected as the performance evaluation index, and its calculation is shown in Equation (21) (21) In the formula, represents the total number of track pairs that should be correctly associated in a scenario, represents the total number of correctly associated track pairs.

[0107] The correct association rate is selected as the comparative performance index. The state-dependent transition probability interactive multi-model interrupted track association (IMMSDP-TSA) algorithm, the track segment association algorithm based on the prior multi-hypothesis model (TSA-MHPI) algorithm, and the interrupted track association algorithm based on the metric neural network (TSADCNN) algorithm are selected as the comparison objects. The performance comparison experiment is carried out using the trained model and the above three algorithms, and the performance of each algorithm is compared under different numbers of targets and different interruption times.

[0108] Set the interruption duration to 10 s, the radar angle measurement error to , and the ranging error to . The influence of different numbers of targets on the performance of different TSA algorithms is shown in Table 3. The method in this embodiment always maintains a high correct association rate as the number of tracks increases.

[0109] Table 3: Correct association rates of each algorithm under different numbers of tracks

[0110] Set the number of targets to 500, and the radar angle measurement error to , the ranging error is , the influence of different interruption times on the performance of different TSA algorithms is shown in Table 4.

[0111] Table 4: The correct association rate of each algorithm at different interruption times

[0112] As described in the above table, the increase of the interruption duration and the number of tracks will increase the association difficulty of each algorithm. Among them, the algorithm of this embodiment can maintain good performance in most cases.

[0113] In the above text, an embodiment of a method for associating interrupted tracks based on track prediction has been described in detail. Based on the method for associating interrupted tracks based on track prediction described in the above embodiment, the embodiment of the present invention also provides a system for associating interrupted tracks based on track prediction corresponding to this method.

[0114] Figure 10 FIG. is a schematic block diagram of a system for associating interrupted tracks based on track prediction provided by an embodiment of the present invention. In this embodiment, the system 1000 for associating interrupted tracks based on track prediction can be divided into multiple functional units according to the functions it performs. The unit referred to in the present invention means a series of computer program segments that can be executed by at least one processor and can complete fixed functions, and are stored in the memory.

[0115] The rough track association unit 1001 is used to perform rough association on all old tracks and new tracks in the scene based on time matching and speed matching, and assign values to the association matrix according to the rough association results.

[0116] The track data normalization processing unit 1002 is used to perform normalization processing on the track data of all track pairs after rough association.

[0117] The track data preprocessing unit 1003 is used to preprocess the track data of old tracks and new tracks, and the preprocessing includes mean filtering and variable dimension expansion operations.

[0118] The track data prediction unit 1004 is used to obtain the track data of the predicted new track based on the preprocessed old track data through a pre-trained track prediction model based on time series.

[0119] The association classification unit 1005 is used to obtain the association classification result of the track pair through a pre-trained time series classification model based on the preprocessed old track data, new track data, and predicted new track data.

[0120] The association result acquisition unit 1006 is used to reassign values to the association matrix according to the association classification result, and obtain the interrupted track association result based on the reassigned association matrix.

[0121] The interrupted track association system based on track prediction in this embodiment is used to implement the aforementioned interrupted track association method based on track prediction. Therefore, the specific implementation in this system can be seen in the embodiment part of the interrupted track association method based on track prediction in the previous text. Therefore, its specific implementation can refer to the descriptions of the corresponding individual embodiment parts and will not be elaborated here.

[0122] In addition, since the interrupted track association system based on track prediction in this embodiment is used to implement the aforementioned interrupted track association method based on track prediction, its function corresponds to that of the above method and will not be elaborated here.

[0123] Figure 11 FIG. 1100 is a schematic structural diagram of a terminal 1100 provided by an embodiment of the present invention, including: a processor 1110, a memory 1120, and a communication unit 1130. When the processor 1110 is used to implement the interrupted track association program based on track prediction stored in the memory 320, the following steps are implemented: Perform rough association on all old tracks and new tracks in the scene based on time matching and speed matching, and assign values to the association matrix according to the rough association results; Perform normalization processing on the track data of all track pairs after rough association; Perform preprocessing on the track data of old tracks and new tracks, and the preprocessing includes mean filtering and variable dimension expansion operations; Based on the preprocessed old track data, obtain the track data of the predicted new track through a pre-trained track prediction model based on time series; Based on the preprocessed old track data, new track data, and predicted new track data, obtain the association classification result of the track pair through a pre-trained time series classification model; Reassign values to the association matrix according to the association classification result, and obtain the interrupted track association result based on the reassigned association matrix.

[0124] The present invention also provides a computer storage medium, and the storage medium here may be a magnetic disk, an optical disk, a read-only memory (abbreviation: ROM), a random access memory (abbreviation: RAM), etc.

[0125] The computer storage medium stores an interrupted track association program based on track prediction. When the interrupted track association program based on track prediction is executed by a processor, the following steps are implemented: Perform rough association on all old tracks and new tracks in the scene based on time matching and speed matching, and assign values to the association matrix according to the rough association results; Normalize the track data of all track pairs after rough association; Preprocess the track data of the old track and the new track, and the preprocessing includes mean filtering and variable dimension expansion operations; Based on the preprocessed old track data, obtain the track data for predicting the new track through a pre-trained track prediction model based on time series; Based on the preprocessed old track data, new track data, and predicted new track data, obtain the association classification result of the track pair through a pre-trained time series classification model; Reassign the association matrix according to the association classification result, and obtain the interrupted track association result based on the reassigned association matrix.

[0126] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be obvious to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to these embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for associating interrupted tracks based on track prediction, characterized in that: The following steps are involved: Based on time matching and speed matching, all old and new tracks in the scene are roughly associated, and the association matrix is ​​assigned values ​​according to the rough association results; Normalize the track data of all track pairs after rough association; Preprocess the track data of the old track and the new track, including mean filtering and variable dimension expansion operations; Based on the pre-processed old track data, the track data for predicting the new track is obtained through the pre-trained time series-based track prediction model; Based on the pre-processed old track data and new track data, and the predicted new track data, the associated classification results of the track pairs are obtained through the pre-trained time series classification model; The association matrix is ​​reassigned according to the association classification result, and the interrupted track association result is obtained based on the reassigned association matrix.

2. The interrupted track association method based on track prediction according to claim 1 is characterized in that: Based on the preprocessed old track data, the track data for predicting the new track is obtained through the pre-trained time series-based track prediction model, including: The first linear layer transforms the dimension of the input data; the input data refers to the preprocessed old track data; The output data of the first linear layer flows into the Mamba model and the reverse Mamba model respectively, processing the sequence data from the forward and reverse directions; The output data of the Mamba model and the reverse Mamba model are combined by an addition operation; The merged data is normalized through the first normalization layer; Perform nonlinear feature extraction on the output data of the first normalization layer through the feedforward network layer; The output data of the first normalization layer and the feedforward network layer are added and then input into the second normalization layer for normalization; The output data of the second normalization layer is input into the second linear layer, and the second linear layer outputs the track data for predicting the new track.

3. The interrupted track association method based on track prediction according to claim 1 or 2, characterized in that: The objective function of the track prediction model in the training phase is to use contrast loss to calculate the loss value between the track data of the predicted new track and the actual new track, which is defined as: in, is a binary label. If the old track and the actual new track in the sample come from the same target, then ,otherwise ; represents the separation distance, which is the shortest distance between two track segments belonging to different targets in the sample; Represents the Euclidean distance between the predicted new track and the actual new track.

4. The interrupted track association method based on track prediction according to claim 3 is characterized in that: Based on the pre-processed old track data and new track data, and the predicted new track data, the associated classification results of the track pairs are obtained through the pre-trained time series classification model, including: Perform subtraction operation on the preprocessed old track data and the new track data; Perform a splicing operation on the subtraction result, the new track data, and the predicted new track data; The splicing results are input into a pre-trained time series classification model, and the time series classification model outputs the associated classification results of the track pairs.

5. The interrupted track association method based on track prediction according to claim 4 is characterized in that: The objective function of the time series classification model in the training phase is to use the cross entropy loss function to calculate the loss value between the predicted probability and the true distribution. The output of the time series classification model for each sample is , is the probability that the old track and the new track do not belong to the same target, is the probability that the old track and the new track belong to the same target, and the true label is 0 or 1. The corresponding cross entropy loss function is expressed as: in is the sample size; The total loss function is expressed as: .

6. The interrupted track association method based on track prediction according to claim 5 is characterized in that: The splicing results are input into the pre-trained time series classification model, and the time series classification model outputs the associated classification results of the track pairs, including: The splicing results are respectively fed into the increase filter, the reduction filter, the peak filter, and the first Inception module; The output data of the increase filter, the reduction filter, the peak filter, and the first Inception module are concatenated to obtain the integrated features; The integrated features are passed through the second Inception module twice to extract multi-scale features; Perform the first residual connection between the input data and the extracted multi-scale features; The result of the first residual connection passes through the third Inception module three times to extract multi-scale features again; Perform a second residual connection on the result of the second residual connection and the multi-scale features extracted again; The second residual connection result is passed through a global average pooling layer to reduce the feature dimension; The output data of the global average pooling layer is passed through the linear layer to output the associated classification results of the track pair.

7. The interrupted track association method based on track prediction according to claim 1 is characterized in that: The association matrix is ​​reassigned according to the association classification result, and the interrupted track association result is obtained based on the reassigned association matrix, specifically including: The associated classification results include Old track and New tracks Probability of not belonging to the same target The probability that and belong to the same target ; In the correlation matrix Line The elements of the column are Old track and New tracks The association mark between ; It means the two are not related. Indicates that the two are related; when And the current association tag When the associated tag Update to 0; Defining a Collection It is initialized to empty and used to store the serial numbers of the new and old tracks that are finally associated; For the incidence matrix that satisfies Traverse the track pairs and find the center of each track pair The old track number with the largest value With new tracks , marking it as ; Mark Add to Collection and the first Row and Set all elements of the column to 0, and then recalculate the occurrence matrix that satisfies Traverse the track pairs; the final set The content in is the interrupted track association result.

8. An interrupted track association system based on track prediction, characterized in that: include: The track coarse association unit is used to roughly associate all old tracks and new tracks in the scene based on time matching and speed matching, and assign values ​​to the association matrix according to the coarse association results; A track data normalization processing unit, used for normalizing the track data of all track pairs after rough association; A track data preprocessing unit is used to preprocess the track data of the old track and the new track, and the preprocessing includes mean filtering and variable dimension expansion operations; A track data prediction unit, used for obtaining track data for predicting a new track based on the pre-processed old track data through a pre-trained track prediction model based on a time series; An association classification unit is used to obtain association classification results of track pairs based on the pre-processed old track data and new track data and predict new track data through a pre-trained time series classification model; The association result acquisition unit is used to re-assign the association matrix according to the association classification result to obtain the interrupted track association result.

9. A terminal, characterized in that: include: A memory, used for storing an interrupted track association program based on track prediction; A processor is used to implement the steps of the method for interrupting track association based on track prediction as described in any one of claims 1 to 7 when executing the interrupting track association program based on track prediction.

10. A computer-readable storage medium, characterized in that: The readable storage medium stores an interrupted track association program based on track prediction, and when the interrupted track association program based on track prediction is executed by a processor, the steps of the interrupted track association method based on track prediction as described in any one of claims 1 to 7 are implemented.

Images