Multi-target resolution method based on track space-time characteristics and dynamic similarity measure

Abstract

This invention discloses a multi-target discrimination method based on the spatiotemporal features and dynamic similarity measurement of flight tracks, comprising the following steps: acquiring raw flight track data; preprocessing the raw flight track data to generate multiple continuous flight track segments; extracting spatiotemporal features and statistical features to construct a similarity matrix; performing hierarchical flight track clustering analysis on the similarity matrix, the result of which is n clusters divided according to the similarity matrix; loading a relational model, performing conflict resolution and target identification, and calculating the flight track association confidence score; and generating a flight track set based on the flight track association confidence score. According to the above technical solution, multi-target echo signals in radar seeker tracking dense target scenarios can be distinguished, achieving high-accuracy target flight track association and high-precision low-latency processing.

Description

Technical Field

[0001] This invention relates to the field of multi-target tracking and recognition technology, and more specifically, to a multi-target discrimination method based on the spatiotemporal characteristics of flight tracks and dynamic similarity measurement. Background Technology

[0002] In scenarios with multiple target tracks, such as radar, radio sensor networks, autonomous driving, and drone swarms, multi-target discrimination typically relies on single-frame data association (e.g., nearest neighbor algorithms, Joint Probabilistic Data Association (JDPA)) or filter-based prediction (e.g., Kalman filtering, particle filtering). However, due to the dense target concentration and overlapping tracks in these scenarios, multi-target discrimination often faces the following challenges:

[0003] 1) Targets with similar motion features are difficult to distinguish, resulting in a high error rate in the association of the discrimination results;

[0004] 2) High computational complexity, making it difficult to meet real-time requirements;

[0005] 3) It is sensitive to noise, and its performance degrades significantly in environments with low signal-to-noise ratio.

[0006] To address these issues, existing technologies typically rely on globally optimal Hungarian algorithms or deep learning models. While these methods partially solve the problems, the Hungarian algorithm is computationally intensive, and deep learning models depend on large amounts of labeled data and have poor generalization capabilities. Therefore, there is an urgent need for an efficient, robust, and computationally manageable method for track association and discrimination to solve the target confusion problem caused by dense targets, intersecting tracks, or noise interference in complex environments. Summary of the Invention

[0007] To achieve the above objectives, this application provides a multi-target discrimination method based on track spatiotemporal features and dynamic similarity measurement, comprising the following steps:

[0008] Acquire raw track data, preprocess the raw track data, and generate multiple continuous track segments;

[0009] Spatiotemporal and statistical features are extracted from continuous track segments to construct a similarity matrix and quantify the spatiotemporal similarity between tracks;

[0010] Hierarchical track clustering analysis is performed on the similarity matrix, and the result of the hierarchical track clustering analysis is n clusters divided according to the similarity matrix;

[0011] Load the relational model, perform conflict resolution and target identification through the relational model, and calculate the track association confidence.

[0012] A set of tracks is generated based on the track association confidence.

[0013] The original track data consists of the original track point set collected by the sensors; the preprocessing includes: performing time alignment, noise reduction filtering and interpolation completion on the original track data.

[0014] Furthermore, the spatiotemporal characteristics include direction of motion, rate of change of velocity, variance of acceleration, and consistency of heading angle; the statistical characteristics include: track length, duration, and position distribution entropy.

[0015] The similarity matrix is ​​constructed by calculating the Hausdorff distance between consecutive track segments, and the calculation method is as follows:

[0016]

[0017] Where a and b are waypoints, A and B are the sets of waypoints, and d(a,b) is the Euclidean distance between points a and b, i.e., d(a,b) = ||ab||.

[0018] Hierarchical track clustering analysis includes:

[0019] Based on the similarity matrix, the improved DBSCAN algorithm is used to perform initial clustering of the tracks, grouping tracks with high similarity into the same cluster;

[0020] Perform conflict detection on the initial clustering results;

[0021] Conflict detection refers to performing spatiotemporal conflict analysis on tracks within each cluster. If there are spatiotemporally conflicting tracks within the same cluster, sub-clustering is initiated.

[0022] Specifically, spatiotemporal conflicts include overlapping positions, large differences in speed, and large differences in track length;

[0023] During the collision detection, if two tracks within a cluster are detected to overlap at time t but with opposite velocity directions, and the collision score is greater than a set threshold, then the initiator clustering is initiated.

[0024] The conflict score is calculated as follows: Score = α·Pc + β·Sc + γ·Ac, where α, β, and γ are weight coefficients that can be optimized through training; Pc is the positional overlap; Sc is the velocity difference; and Ac is the acceleration difference.

[0025] Sub-clustering refers to: based on the improved DBSCAN algorithm, reducing the neighborhood radius to separate dense conflict tracks, and then re-dividing the clusters based on conflict characteristics;

[0026] When starting the sub-clustering, change the similarity weights.

[0027] Furthermore, the relation model adopts a bipartite graph model structure, where the node set of the relation model is the track to be associated, and the edge weight is the similarity score of the track.

[0028] The relational model uses the Hungarian algorithm to solve for maximum weight matching, with added constraints; the constraints include: maximum association delay and motion continuity.

[0029] Define low-confidence matching detection rules, initiate a backtracking mechanism for low-confidence association results, and dynamically correct the association relationship based on historical data;

[0030] The formula for low-confidence matching detection is as follows:

[0031] If conf ij If the match is less than the set detection threshold, it is marked as a low-confidence match.

[0032] Furthermore, conflict resolution through a relational model includes:

[0033] For tracks that cannot be uniquely associated, target attribute information is fused to assist in decision-making; wherein, the target attribute information includes: radar cross-section (RCS), optical size features, and texture entropy features;

[0034] Update the target identity probability based on Bayesian inference;

[0035] The identity determination rules of the Hungarian algorithm are used to calculate the confidence level of track association.

[0036] determination Whether it is within the confidence threshold; for association pairs with a confidence level below the threshold, the target attribute information is used for verification, and the independent target track is output.

[0037] The multi-target discrimination method provided by this invention offers a new approach to multi-target discrimination in dense target scenarios from the perspective of signal processing technology. It distinguishes the echo signals of multiple targets in dense target scenarios when the radar seeker is tracking them, and achieves high-accuracy correlation and high-precision low-latency processing of target tracks. Attached Figure Description

[0038] Figure 1 This is a schematic diagram of the steps of a multi-target resolution method according to an embodiment of the present invention. Detailed Implementation

[0039] This invention proposes a multi-target discrimination method based on the spatiotemporal features of the track and dynamic similarity measurement. The DBSCAN algorithm is used to classify and identify the track features of multiple radar targets. The spatiotemporal features, statistical characteristics and contextual information are fused. A conflict resolution mechanism is used to deal with target intersection or occlusion scenarios, so as to achieve high-precision and low-complexity track association and thus target recognition.

[0040] The specific implementation of the present invention will now be described in detail with reference to the accompanying drawings.

[0041] The multi-target resolution method provided by this invention is as follows: Figure 1 As shown, it includes the following steps:

[0042] Step S100: Obtain raw track data;

[0043] The raw track data is the raw track point set collected by sensors such as radar detection;

[0044] Step S110: Preprocess the original track data to generate multiple continuous track segments;

[0045] The original track data is time aligned and denoised; then, missing data is filled by interpolation (such as cubic spline interpolation) to generate continuous track segments.

[0046] This invention provides a multi-target resolution case for radar seeker tracking multiple targets. After inputting the track points detected by the radar, the input data is time-aligned, denoised, and interpolated to obtain a set of pre-processed continuous track segments, generating 10 track segments.

[0047] Step S120: Extract spatiotemporal features and statistical features from continuous track segments, construct a similarity matrix, and quantify the spatiotemporal similarity between tracks;

[0048] Specifically, spatiotemporal characteristics include direction of motion, rate of change of velocity, variance of acceleration, and consistency of heading angle;

[0049] Based on the spatiotemporal features, statistical features can be extracted, including: track length, duration, and position distribution entropy.

[0050] In constructing the similarity matrix, the Hausdorff distance between consecutive track segments is first calculated.

[0051] Hausdorff distance is used to measure the maximum and minimum spatial deviation between two tracks, and it is calculated as follows:

[0052]

[0053] Where a and b are waypoints, A and B are the sets of waypoints, and d(a,b) is the Euclidean distance between points a and b, i.e., d(a,b) = ||ab||.

[0054] The similarity matrix is ​​subjected to Laplace matrix transformation and dimensionality reduction to quantify the spatiotemporal similarity between tracks.

[0055] In the case provided by the present invention, the preprocessed track is extracted to extract the track direction, velocity curve, acceleration variance, heading angle consistency, track length, duration, and position distribution entropy.

[0056] After obtaining the spatiotemporal features of the flight path, subsequent steps can be performed to analyze the dynamic similarity of multiple targets.

[0057] Step S130: Perform hierarchical track clustering analysis on the similarity matrix. The result of the clustering analysis is n clusters divided according to the similarity matrix.

[0058] Hierarchical track clustering analysis includes:

[0059] 1) Based on the similarity matrix, the improved DBSCAN algorithm is used to perform initial clustering of the tracks, grouping tracks with high similarity into the same cluster; the operation of the DBSCAN algorithm includes: introducing an adaptive neighborhood radius and dynamically adjusting the threshold parameter of the clustering algorithm cost function according to the track density;

[0060] 2) In scenarios with dense targets and intersecting tracks, targets with similar motion characteristics are difficult to distinguish, leading to inaccuracies in the initial clustering results. Therefore, this invention proposes conflict detection for the initial clustering results: spatiotemporal conflict analysis is performed on tracks within each cluster. If spatiotemporally conflicting tracks exist within the same cluster, sub-clustering is initiated. Spatiotemporal conflicts specifically include situations such as overlapping positions, large velocity differences, and large track length differences. During conflict detection, if two tracks within a cluster are detected to overlap at time t but with opposite velocity directions, and the conflict score is greater than a set threshold, then sub-clustering is initiated.

[0061] The conflict score is calculated as follows: Score = α·Pc + β·Sc + γ·Ac, where α, β, and γ are weight coefficients that can be optimized through training; Pc is the positional overlap; Sc is the velocity difference; and Ac is the acceleration difference.

[0062] Sub-clustering refers to: based on an improved DBSCAN algorithm, reducing the neighborhood radius to separate densely conflicting tracks, and then re-dividing clusters based on conflict features. When initiating sub-clustering, the similarity weights are changed.

[0063] After dividing the data into multiple clusters based on the similarity matrix, the track can be identified and the target determined based on the relational model. The first step is to define the relational model.

[0064] In the case provided by this invention, the Hausdorff distance similarity matrix of each track is calculated, the similarity matrix is ​​transformed by Laplace matrix and the dimension is reduced, and then the track set of the similarity matrix is ​​divided into 3 clusters by introducing the improved DBSCAN algorithm with adaptive neighborhood radius.

[0065] Next, the relationships between tracks within the cluster are identified using a relational model.

[0066] Step S140: Define the relational model;

[0067] 1) In terms of structure: The relational model adopts a bipartite graph model structure, where the node set is the track to be associated, and the edge weight of the bipartite graph model is the similarity score of the track.

[0068] In this step, a method for calculating track similarity based on kinematic features, statistical features, and dynamic time is used to simplify the calculation rules and reduce the complexity of the calculation.

[0069] The method for calculating track similarity is: s(u i ,v j )=α q ·s dtw +β q ·s kin +γ q ·s stst ,

[0070] Where, α q β q γ q U is the weighting coefficient. i For historical navigation, v j For new detection tracks, s dtw For dynamic time warping score, s kin For kinematic feature scores, s stat The score is a statistical feature score.

[0071] 2) Regarding rule determination: The Hungarian algorithm is introduced to solve the maximum weight matching, and constraints are added, such as maximum associated delay and motion continuity.

[0072] 3) Further optimization of the relationship model: Define low-confidence matching detection rules, initiate a backtracking mechanism for low-confidence association results, and dynamically correct the association relationship based on historical data:

[0073] The formula for low-confidence match detection is as follows: If conf ij If the match is less than the set detection threshold, it is marked as a low-confidence match.

[0074] After the relational model is constructed and optimized, step S141 can be executed:

[0075] Step S141: Load the relational model and perform conflict resolution and target identification through the relational model;

[0076] First, locate tracks that cannot be uniquely associated, fuse target attribute information, and make auxiliary decisions; among which, target attribute information generally includes: radar cross-section (RCS), optical size characteristics, and texture entropy characteristics;

[0077] Secondly, the target identity probability is updated based on Bayesian inference;

[0078] Specifically, the Bayesian update formula is as follows:

[0079] Among them, ID k Let f be the discrete target label, and f be the feature vector corresponding to the target attribute information; p(ID) k Prior probability, i.e., the confidence level of identity based on historical trajectory association results, p(f|ID) k ) is the probability density function, in the identity ID k The probability of observing feature f under given conditions; p(ID) k |f) Posterior probability, given observed features f, the target identity is ID k The update probability; M is the total number of candidate identities.

[0080] The total number of candidate identities is determined by the module that performs conflict detection on the clustering results.

[0081] Furthermore, the probability density function p(f|ID) k The calculation method for ) is as follows:

[0082] p(f|ID k )=p(f RCS ID k )·p(f size ID k )·p(f texture ID k ), where f RCS For radar RCS characteristics, f size For optical size characteristics, f texture This is a texture entropy feature.

[0083] Prior probability p(ID) k The update method for p is: new (ID k )==η·p(ID k |f)+(1-η)·p prev (ID k ), where η is the learning rate, which controls the weight of new observations;

[0084] Next, the track association confidence score is calculated using the identity determination rules of the Hungarian algorithm:

[0085] Step S150: Generate a set of tracks based on the track association confidence.

[0086] In this step, the determination is made. Whether it is within the confidence threshold; for association pairs with a confidence level below the threshold, RCS feature verification is used to output independent target tracks.

[0087] In the case provided by this invention, the Hungarian algorithm is used to associate tracks. For the associated pairs with confidence levels below a threshold, the RCS feature is called for verification, and the eight independent target tracks are output after resolution.

[0088] Tests have demonstrated that the multi-target discrimination method provided by this invention improves the correlation accuracy by ≥15% and reduces the computational complexity to 0 in dense target scenarios, meeting real-time processing requirements. At the same time, it supports multi-modal sensor data fusion, enhances adaptability to complex environments, effectively improves the classification and identification of dense multi-targets, and thus improves the accuracy of radar-guided multi-target detection and tracking.

[0089] The multi-target discrimination method based on the spatiotemporal characteristics and dynamic similarity measurement of the track has shown good results. Researching a multi-target discrimination method based on the spatiotemporal characteristics and dynamic similarity measurement of the track is of great significance and value. It lays a theoretical foundation for radar seeker detection and tracking of dense multi-targets in subsequent engineering applications, and has important theoretical research significance and engineering application value.

[0090] The above-disclosed embodiments are merely a few specific examples of the present invention. However, the present invention is not limited thereto, and any variations that can be conceived by those skilled in the art should fall within the protection scope of the present invention.

Claims

1. A multi-target discrimination method based on trajectory spatiotemporal features and dynamic similarity measurement, characterized in that, Includes the following steps: Acquire raw track data, preprocess the raw track data, and generate multiple continuous track segments; Spatiotemporal and statistical features are extracted from the continuous track segments to construct a similarity matrix and quantify the spatiotemporal similarity between tracks; Hierarchical track clustering analysis is performed on the similarity matrix, and the result of the hierarchical track clustering analysis is n clusters divided according to the similarity matrix; Load the relational model, perform conflict resolution and target identification through the relational model, and calculate the track association confidence. Based on the track association confidence, a track set is generated; The relationship model adopts a bipartite graph model structure, where the node set of the relationship model is the track to be associated, and the edge weight is the similarity score of the track. The relational model uses the Hungarian algorithm to solve for maximum weight matching, with added constraints; these constraints include: maximum association delay and motion continuity. Define low-confidence matching detection rules, initiate a backtracking mechanism for low-confidence association results, and dynamically correct the association relationship based on historical data; The low-confidence matching detection formula is as follows: , like If the match is less than the set detection threshold, it is marked as a low-confidence match; The conflict resolution performed through the relational model includes: For tracks that cannot be uniquely associated, target attribute information is fused to assist in decision-making; wherein, the target attribute information includes: radar cross-section (RCS), optical size features, and texture entropy features; The target identity probability is updated based on Bayesian inference; where the Bayesian formula is: ,in, For discrete target labels, The feature vector corresponding to the target attribute information. For prior probability, Let be the probability density function. For posterior probability, The total number of candidate identities; The track association confidence score is calculated using the identity determination rules of the Hungarian algorithm and is expressed as follows: .

2. The multi-target resolution method according to claim 1, characterized in that, The original track data is the set of original track points collected by the sensors; The preprocessing includes performing time alignment, noise reduction filtering, and interpolation completion on the original track data.

3. The multi-target resolution method according to claim 1, characterized in that, The spatiotemporal features include motion direction, rate of change of velocity, acceleration variance, and heading angle consistency; The statistical features include: track length, duration, and position distribution entropy.

4. The multi-target resolution method according to claim 3, characterized in that, The similarity matrix is ​​constructed by calculating the Hausdorff distance between consecutive track segments, and the calculation method is as follows: ， in, , For waypoints, , For the set of track points, for , The Euclidean distance between two points, i.e. .

5. The multi-target resolution method according to claim 1, characterized in that, The hierarchical trajectory clustering analysis includes: Based on the similarity matrix, the improved DBSCAN algorithm is used to perform initial clustering of the tracks, grouping tracks with high similarity into the same cluster; Conflict detection is performed on the initial clustering results; The conflict detection refers to: performing spatiotemporal conflict analysis on the tracks within each cluster, and initiating sub-clustering if there are spatiotemporally conflicting tracks within the same cluster; The spatiotemporal conflicts include overlapping positions, large differences in speed, and large differences in track length. During the collision detection, if two tracks within a cluster are detected to overlap at time t but with opposite velocity directions, and the collision score is greater than a set threshold, then the initiator clustering is initiated.

6. The multi-target resolution method according to claim 5, characterized in that, The conflict score is calculated as follows: ,in, These are weighting coefficients, which can be optimized through training; This refers to the degree of positional overlap. For speed differences; This is due to differences in acceleration.

7. The multi-target resolution method according to claim 5, characterized in that, The sub-clustering refers to: based on the improved DBSCAN algorithm, reducing the neighborhood radius to separate dense conflict tracks, and then re-dividing the clusters based on conflict characteristics; When starting the sub-clustering, change the similarity weights.

8. The multi-target resolution method according to claim 1, characterized in that, determination Whether it is within the confidence threshold; for association pairs with a confidence level below the threshold, the target attribute information is used for verification, and the independent target track is output.

Images