Low-altitude radar multi-target secondary tracking method based on trajectory level correlation and robust features

Abstract

The invention discloses a low-altitude radar multi-target secondary tracking method based on trajectory level association and robust features, relates to the field of low-altitude safety, and is a low, small and slow target tracking method based on multi-dimensional robust feature association and spatio-temporal index pruning. According to the method, by introducing a decoupling architecture of'primary tracking extraction state and secondary tracking maintenance logic ', a tracking task is divided into front-section state estimation and rear-end track management, the front-section focuses on high-sensitivity state extraction, and the rear-end focuses on long-time-sequence track maintenance. According to the method, the problem of track fragmentation caused by the fact that the target enters the Doppler blind area or is shielded under the low-altitude complex background is solved.

Description

Technical Field

[0001] This invention relates to the field of low-altitude safety, and in particular to a low-altitude radar multi-target secondary tracking method based on trajectory-level correlation and robust features. Background Technology

[0002] In complex low-altitude scenarios, ground obstruction, multipath propagation, and clutter interference cause intermittent target measurements, resulting in frequent track breaks, false track mixing, and tag (ID) switching in the output track of primary trackers (regardless of whether JPDA / MHT or random finite set-based filters are used). While directly increasing the survival probability to extend the track can alleviate the breaks, it leads to the accumulation of false targets and increased computational overhead. On the other hand, traditional point-based data association is susceptible to single-frame measurement fluctuations under high clutter conditions, making it difficult to stably distinguish between real targets and accompanying clutter. Therefore, a secondary tracking mechanism tailored to the characteristics of engineering data is needed. This mechanism should utilize temporal information and robust features at the trajectory level for re-association and repair, along with efficient pruning to ensure real-time performance. Summary of the Invention

[0003] In view of the above-mentioned deficiencies of the prior art, the technical problem to be solved by the present invention is how to meet the application requirements of real-time performance.

[0004] To achieve the above objectives, this invention provides a low-altitude radar multi-target secondary tracking method based on trajectory-level correlation and robust features. The method is characterized by being a low-altitude, small, and slow target tracking method based on multidimensional robust feature correlation and spatiotemporal index pruning. This method introduces a decoupled architecture of "first-stage tracking to extract state and second-stage tracking to maintain logic", dividing the tracking task into front-end state estimation and back-end track management. The front-end focuses on high-sensitivity state extraction, while the back-end focuses on long-term track maintenance.

[0005] Furthermore, the method includes three core technical modules: extraction of multidimensional robust micro-feature space, trajectory-level correlation measurement and interpolation repair mechanism, and efficient pruning based on spatiotemporal indexing and feature scoring.

[0006] Furthermore, in the extraction module of the multidimensional robust micro-feature space, the secondary tracking module does not directly process the original measurement points, but performs deep motion morphology analysis on the track fragment sequence output by the primary tracker; in view of the difference in spatiotemporal distribution between low, small, slow targets and environmental clutter, the method extracts the following five-dimensional robust features, namely radial velocity consistency feature, motion smoothness feature, trajectory meandering feature, time accumulation feature, and average signal-to-noise ratio feature.

[0007] Furthermore, the radial velocity consistency characteristic refers to the calculation of the average radial velocity of each point within the track segment using radar Doppler measurement information. The radial velocity change of a real target is usually smooth and continuous, while the radial velocity of ground clutter usually approaches zero or exhibits disordered random changes. The motion smoothness feature refers to evaluating the physical coherence of the trajectory by calculating the cosine of the angle between velocity vectors at adjacent moments. The linear acceleration and angular velocity changes of a real flying target are constrained by dynamics, while false tracks formed by clutter associations are often accompanied by non-physical, violent back-and-forth jumps. The trajectory meandering degree is defined as the ratio of the actual path length of the trajectory to the straight-line distance between the beginning and end; The method uses this feature to identify typical clutter trajectories that “spin in place.” Although such trajectories may have a high signal-to-noise ratio, their effective displacement is extremely small and their meandering is much greater than that of real aircraft. The aforementioned time accumulation feature refers to the ability to distinguish between transient noise flickering and persistent real targets by utilizing the cumulative effect of the time dimension through normalized track survival time. The average signal-to-noise ratio (SNR) characteristic refers to the SNR distribution within the statistical track segment. The echo energy of the real target is relatively stable in a short period of time, while the SNR of the interference signal fluctuates drastically and is often at a low level.

[0008] Furthermore, the trajectory-level correlation measurement and interpolation repair mechanism proposes a coarse-to-fine correlation determination process for fragmented tracks that have been divided into multiple segments: Step 1: Comprehensive correlation cost modeling, constructing a comprehensive function that includes motion direction connection cost, radial velocity difference cost, signal strength similarity cost, and spatial Euclidean distance cost. This invention enhances the distinguishability of targets in intersecting paths and dense clutter areas by increasing the weight of the velocity dimension. Step 2: Bidirectional optimal matching logic. To avoid local optima caused by greedy association, a logic combining forward matching and reverse confirmation is adopted. Only track pairs that satisfy bidirectional uniqueness and have a cost value lower than the dynamic threshold are judged as the same target. Step 3, Cubic Spline Interpolation Smoothing: For two track segments that are determined to be successfully associated, instead of using a simple straight line connection, the motion vector before the vanishing point is used as the control point, and the coordinates during the occlusion period are completed by cubic spline interpolation; this ensures that the stitched trajectory remains continuous and smooth in the dimensions of position, velocity and even acceleration.

[0009] Furthermore, to prevent the historical track buffer pool from causing the "curse of dimensionality" as the exploration time increases, the efficient pruning based on spatiotemporal indexing and feature scoring implements multi-level optimizations oriented towards engineering real-time performance, including hash grid spatial indexing and three-state scoring lifecycle management.

[0010] Furthermore, the hash grid spatial index discretizes the monitoring area into a uniform grid, and historical tracks are mapped to the corresponding grid according to their ending positions; new tracks only retrieve candidate objects within the local neighborhood covered by their motion gates, reducing the complexity of the association search from Reduced to approximately constant level .

[0011] Furthermore, the three-state scoring lifecycle management module constructs a comprehensive scoring function based on the aforementioned five-dimensional features, dividing the track into candidate state, confirmed state, and extinguished state. Tracks with scores below the threshold are immediately eliminated, effectively suppressing the dwell time of false tracks in the system.

[0012] Furthermore, the method includes the following steps: Step S1: Track segment preprocessing and filtering. Obtain discrete track sequences from a single tracker, extract state estimation vectors containing spatial position and three-dimensional velocity, and simultaneously obtain the corresponding signal-to-noise ratio distribution information. Step S2: Five-dimensional robust micro-feature extraction. For each segment of the trajectory, calculate its radial velocity fluctuation variance, mean cosine of the velocity angle between adjacent time points, path length to displacement ratio, number of live frames, and average SNR to construct the fingerprint feature vector of the trajectory. Step S3: Spatiotemporal hash index construction. Establish a spatial hash table based on the monitoring area, store the termination position of historical broken tracks as key values, and when a new track starts, calculate its search radius based on the maximum speed constraint. Only query historical tracks within the hash grid covered by this radius to achieve sparsity of the associated space. Step S4: Track-level bidirectional matching decision, calculate the directional connection cost, radial velocity difference cost and spatial distance cost between the new and old tracks; use the bidirectional matching algorithm, that is, the best object of the new track must also be regarded as the best object, to determine the final association pair, effectively preventing ID error exchange under cross paths. Step S5, Repair and Smoothing of Broken Gaps: For the two track segments that are determined to be related, the historical ID is inherited; for the missing observation frames in the middle, the cubic spline interpolation function $S(t)$ is used for reconstruction to ensure that the trajectory does not produce instantaneous velocity changes at the breakpoint. Step S6, Differentiated Pruning and Lifecycle Management: Calculate the Overall Quality Score For tracks with high scores, their retention period in the buffer pool is extended to cope with long-term occlusion; for tracks with extremely low scores or exhibiting non-physical motion characteristics, they are identified as clutter and memory resources are released immediately to ensure that the single-frame association time is maintained at the millisecond level.

[0013] Further, in step S1, the complex-valued signal output by the radar is processed to obtain the slant range, azimuth, pitch and Doppler velocity, and then input into a primary tracker for real-time state estimation, and outputs a track segment with an initial label.

[0014] The present invention has the following technical effects: (1) Actual test data shows that compared with the first tracking, the second tracking increases the average tracking time from 9.1 frames to 41.1 frames, significantly reducing the generation of track debris.

[0015] (2) The number of times the main target's label was switched was reduced significantly from 64 times to 4 times, achieving stable label locking of the target over a long period of time and across occlusion.

[0016] (3) By eliminating false tracks through morphological features (such as meandering), the number of false tracks in batches is reduced by more than 97.9%, which greatly improves the radar situation display interface.

[0017] The following will further explain the concept, specific structure, and technical effects of the present invention in conjunction with the accompanying drawings, so as to fully understand the purpose, features, and effects of the present invention. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall process of secondary tracking according to a preferred embodiment of the present invention; Figure 2 This is a schematic diagram of trajectory-level association and lifecycle management according to a preferred embodiment of the present invention; Figure 3 This is a schematic diagram of a pruning strategy based on spatiotemporal indexing and Q_score, which is a preferred embodiment of the present invention. Detailed Implementation

[0019] The following description, with reference to the accompanying drawings, illustrates several preferred embodiments of the present invention to make its technical content clearer and easier to understand. The present invention can be embodied in many different forms, and the scope of protection of the present invention is not limited to the embodiments mentioned herein.

[0020] In the accompanying drawings, components with the same structure are indicated by the same numerical designation, and components with similar structures or functions are indicated by similar numerical designations. The dimensions and thicknesses of each component shown in the drawings are arbitrary, and the present invention does not limit the dimensions and thicknesses of each component. To make the illustrations clearer, the thickness of some components has been appropriately exaggerated in the drawings.

[0021] like Figures 1-3As shown, to solve the above-mentioned technical problems, the technical solution adopted by this invention is: a low-altitude, small, and slow target tracking method based on multidimensional robust feature association and spatiotemporal index pruning. This method solves the problem of track fragmentation caused by targets entering Doppler blind zones or being occluded in complex low-altitude backgrounds by introducing a decoupled architecture of "first-time tracking to extract state, second-time tracking to maintain logic". The method mainly includes the following core technical modules: 1. Extraction of multidimensional robust micro-feature space The secondary tracking module does not directly process the original measurement points, but instead performs deep motion morphology analysis on the track fragment sequence output by the primary tracker (such as a GLMB filter). Considering the differences in spatiotemporal distribution between low, small, and slow targets and environmental clutter (such as swaying tree canopies and the edges of tall buildings), this invention extracts the following five-dimensional robust features: Radial velocity consistency characteristics: The average radial velocity of each point within the track segment is calculated using radar Doppler measurements. The radial velocity variation of a real target is usually smooth and continuous, while the radial velocity of ground clutter usually approaches zero or exhibits disordered random variations.

[0022] Motion smoothness characteristics: The cosine of the angle between velocity vectors at adjacent time points is calculated to assess the physical consistency of the trajectory. The linear acceleration and angular velocity changes of a real flying target are constrained by dynamics, while false tracks formed by clutter associations are often accompanied by abrupt, non-physical jumps and abrupt changes.

[0023] Trajectory meandering (geometric feature): defined as the ratio of the actual path length of the trajectory to the straight-line distance between the beginning and end.

[0024] This invention utilizes this feature to identify typical clutter trajectories that "spin in place." Although such trajectories may have a high signal-to-noise ratio, their effective displacement is extremely small, and their meandering is much greater than that of real aircraft.

[0025] Time-cumulative characteristics: By normalizing the track survival time, the cumulative effect of the time dimension is used to distinguish between transient noise flickering and persistent real targets.

[0026] Average signal-to-noise ratio characteristics: SNR distribution within the statistical track segment. The energy of the real target echo is relatively stable over a short period of time, while the SNR of the interference signal fluctuates drastically and is often at a low level.

[0027] 2. Trajectory-level correlation measurement and interpolation repair mechanism

[0028] For fragmented flight paths that are divided into multiple segments, this invention proposes a coarse-to-fine correlation determination process: Comprehensive correlation cost modeling: Constructing a comprehensive function that includes motion direction connection cost, radial velocity difference cost, signal strength similarity cost, and spatial Euclidean distance cost. This invention enhances the distinguishability of targets in intersecting paths and dense clutter areas by increasing the weight of the velocity dimension.

[0029] Bidirectional optimal matching logic: To avoid local optima caused by greedy association, a logic combining forward matching and reverse confirmation is adopted. Only track pairs that satisfy bidirectional uniqueness and whose cost is below the dynamic threshold are determined to be the same target.

[0030] Cubic spline interpolation smoothing: For two successfully associated track segments, this invention no longer uses a simple straight line connection. Instead, it uses the motion vector before the vanishing point as control points and completes the coordinates during the occlusion period through cubic spline interpolation. This ensures that the stitched trajectory remains continuous and smooth in the dimensions of position, velocity, and even acceleration.

[0031] 3. Efficient pruning based on spatiotemporal indexing and feature scoring

[0032] To prevent the "curse of dimensionality" from occurring in the historical track buffer pool as the detection time increases, this invention implements multi-level optimization oriented towards engineering real-time performance: Hash grid spatial indexing: The monitored area is discretized into a uniform grid, and historical tracks are mapped to the corresponding grid according to their termination positions. New tracks only retrieve candidate objects within the local neighborhood covered by their motion gates, reducing the complexity of correlation search from... Reduced to approximately constant level .

[0033] Three-state scoring lifecycle management: Construct a comprehensive scoring function based on the aforementioned five-dimensional features, divide the track into candidate state, confirmed state and disappearance state, and immediately remove tracks with scores below the threshold (such as abnormal meandering), effectively suppressing the residence time of false tracks in the system.

[0034] The specific technical solution adopted in this invention is as follows: First, the complex-valued signal output from the radar is processed to obtain slant range, azimuth, elevation, and Doppler velocity, which are then input into a primary tracker for real-time state estimation, outputting track segments with initial labels. Subsequently, multi-domain feature analysis is performed on these segments, calculating five robust micro-features: motion smoothness, radial velocity consistency, and trajectory shape characteristics, forming a feature space. In the association phase, a spatiotemporal hash index is used to quickly filter potential matching pairs from the historical buffer pool, and track continuation is completed through a two-way optimal matching judgment logic. Successfully matched tracks undergo ID unification and interpolation smoothing, while unmatched tracks are filtered using a quality scoring system, removing low-confidence clutter tracks and retaining only high-value targets for continuous analysis. This scheme significantly improves track continuity while ensuring the system's real-time processing performance under large-scale data through efficient pruning.

[0035] The specific implementation steps are as follows: S1. Track segment preprocessing and filtering: Obtain discrete track sequences from a single tracker, extract state estimation vectors containing spatial position and three-dimensional velocity, and simultaneously obtain the corresponding signal-to-noise ratio distribution information.

[0036] S2. Five-dimensional robust micro-feature extraction: For each segment of the track, calculate its radial velocity fluctuation variance, mean cosine of the velocity angle between adjacent time points, path length to displacement ratio (winding degree), number of live frames, and average SNR to construct the fingerprint feature vector of the track.

[0037] S3. Spatiotemporal Hash Index Construction: Establish a spatial hash table based on the monitored area. Store the termination positions of historical broken tracks as keys. When a new track begins, calculate its search radius based on the maximum velocity constraint, and only query historical tracks within the hash grid covered by this radius, thus achieving sparsity of the associated space.

[0038] S4. Track-level bidirectional matching decision: Calculates the directional connection cost, radial velocity difference cost, and spatial distance cost between the new and old tracks. A bidirectional matching algorithm (i.e., the best object of the new track must also be considered the best object) is used to determine the final associated pair, effectively preventing incorrect ID swapping under intersecting paths.

[0039] S5. Fault Gap Repair and Smoothing: For two correlated track segments, the historical ID is inherited. For missing observation frames in the middle, a cubic spline interpolation function $S(t)$ is used for reconstruction to ensure that the trajectory does not produce instantaneous velocity changes at the breakpoint.

[0040] S6. Differentiated Pruning and Lifecycle Management: Calculating the Overall Quality Score For tracks with high scores, their retention period in the buffer pool is extended to cope with long-term occlusion; for tracks with extremely low scores or exhibiting non-physical motion characteristics (such as high-frequency jumps in place), they are identified as clutter and memory resources are immediately released to ensure that the single-frame association time is maintained at the millisecond level.

[0041] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A low-altitude radar multi-target secondary tracking method based on trajectory-level correlation and robust features, characterized in that, This method is a low-speed target tracking method based on multidimensional robust feature association and spatiotemporal index pruning. By introducing a decoupled architecture of "first-stage tracking to extract state and second-stage tracking to maintain logic", the tracking task is divided into front-end state estimation and back-end track management. The front-end focuses on high-sensitivity state extraction, while the back-end focuses on long-term track maintenance.

2. The low-altitude radar multi-target secondary tracking method based on trajectory-level correlation and robust features as described in claim 1, characterized in that, The method comprises three core technical modules: extraction of multidimensional robust micro-feature space, trajectory-level correlation measurement and interpolation repair mechanism, and efficient pruning based on spatiotemporal indexing and feature scoring.

3. The low-altitude radar multi-target secondary tracking method based on trajectory-level correlation and robust features as described in claim 2, characterized in that, In the extraction module of the multidimensional robust micro-feature space, the secondary tracking module does not directly process the original measurement points, but performs deep motion morphology analysis on the track fragment sequence output by the primary tracker. In view of the difference in spatiotemporal distribution between low, small, slow targets and environmental clutter, the method extracts the following five robust features, namely radial velocity consistency feature, motion smoothness feature, trajectory meandering feature, time accumulation feature, and average signal-to-noise ratio feature.

4. The low-altitude radar multi-target secondary tracking method based on trajectory-level correlation and robust features as described in claim 3, characterized in that, The radial velocity consistency characteristic refers to the average radial velocity of each point within the track segment calculated using radar Doppler measurement information. The radial velocity change of a real target is usually smooth and continuous, while the radial velocity of ground clutter usually approaches zero or exhibits disordered random changes. The motion smoothness feature refers to evaluating the physical coherence of the trajectory by calculating the cosine of the angle between velocity vectors at adjacent moments. The linear acceleration and angular velocity changes of a real flying target are constrained by dynamics, while false tracks formed by clutter associations are often accompanied by non-physical, violent back-and-forth jumps. The trajectory meandering degree is defined as the ratio of the actual path length of the trajectory to the straight-line distance between the beginning and end; The method uses this feature to identify typical clutter trajectories that "spin in place". Although such trajectories may have a high signal-to-noise ratio, their effective displacement is extremely small and their meandering is much higher than that of real aircraft. The aforementioned time accumulation feature refers to the ability to distinguish between transient noise flickering and persistent real targets by utilizing the cumulative effect of the time dimension through normalized track survival time. The average signal-to-noise ratio characteristic refers to the SNR distribution within the statistical track segment. The echo energy of the real target is relatively stable in a short period of time, while the SNR of the interference signal fluctuates drastically and is often at a low level.

5. The low-altitude radar multi-target secondary tracking method based on trajectory-level correlation and robust features as described in claim 4, characterized in that, The trajectory-level correlation measurement and interpolation repair mechanism proposes a coarse-to-fine correlation determination process for fragmented tracks that have been divided into multiple segments: Step 1: Comprehensive correlation cost modeling, constructing a comprehensive function that includes motion direction connection cost, radial velocity difference cost, signal strength similarity cost, and spatial Euclidean distance cost. This invention enhances the distinguishability of targets in intersecting paths and dense clutter areas by increasing the weight of the velocity dimension. Step 2: Bidirectional optimal matching logic. To avoid local optima caused by greedy association, a logic combining forward matching and reverse confirmation is adopted. Only track pairs that satisfy bidirectional uniqueness and have a cost value lower than the dynamic threshold are judged as the same target. Step 3, Cubic Spline Interpolation Smoothing: For two track segments that are determined to be successfully associated, instead of using a simple straight line connection, the motion vector before the vanishing point is used as the control point, and the coordinates during the occlusion period are completed by cubic spline interpolation; this ensures that the stitched trajectory remains continuous and smooth in the dimensions of position, velocity and even acceleration.

6. The low-altitude radar multi-target secondary tracking method based on trajectory-level correlation and robust features as described in claim 5, characterized in that, To prevent the historical track buffer from causing the "curse of dimensionality" as the exploration time increases, the efficient pruning based on spatiotemporal indexing and feature scoring implements multi-level optimizations oriented towards engineering real-time performance, including hash grid spatial indexing and three-state scoring lifecycle management.

7. The low-altitude radar multi-target secondary tracking method based on trajectory-level correlation and robust features as described in claim 6, characterized in that, The hash grid spatial index discretizes the monitoring area into a uniform grid, and historical tracks are mapped to the corresponding grid according to their termination positions; new tracks only search for candidate objects within the local neighborhood covered by their motion gates, reducing the complexity of the association search from Reduced to approximately constant level .

8. The low-altitude radar multi-target secondary tracking method based on trajectory-level correlation and robust features as described in claim 7, characterized in that, The three-state scoring lifecycle management module constructs a comprehensive scoring function based on the aforementioned five-dimensional features, dividing the track into candidate state, confirmed state, and annihilated state. Tracks with scores below the threshold are immediately removed, effectively suppressing the residence time of false tracks in the system.

9. The low-altitude radar multi-target secondary tracking method based on trajectory-level correlation and robust features as described in claim 8, characterized in that, The method includes the following steps: Step S1: Track segment preprocessing and filtering. Obtain discrete track sequences from a single tracker, extract state estimation vectors containing spatial position and three-dimensional velocity, and simultaneously obtain the corresponding signal-to-noise ratio distribution information. Step S2: Five-dimensional robust micro-feature extraction. For each segment of the trajectory, calculate its radial velocity fluctuation variance, mean cosine of the velocity angle between adjacent time points, path length to displacement ratio, number of live frames, and average SNR to construct the fingerprint feature vector of the trajectory. Step S3: Spatiotemporal hash index construction. Establish a spatial hash table based on the monitoring area, store the termination position of historical broken tracks as key values, and when a new track starts, calculate its search radius based on the maximum speed constraint. Only query historical tracks within the hash grid covered by this radius to achieve sparsity of the associated space. Step S4: Track-level bidirectional matching decision, calculate the directional connection cost, radial velocity difference cost and spatial distance cost between the new and old tracks; use the bidirectional matching algorithm, that is, the best object of the new track must also be regarded as the best object, to determine the final association pair, effectively preventing ID error exchange under cross paths. Step S5, Repair and Smoothing of Broken Gaps: For the two track segments that are determined to be related, the historical ID is inherited; for the missing observation frames in the middle, the cubic spline interpolation function $S(t)$ is used for reconstruction to ensure that the trajectory does not produce instantaneous velocity changes at the breakpoint. Step S6, Differentiated Pruning and Lifecycle Management: Calculate the Overall Quality Score For tracks with high scores, their retention period in the buffer pool is extended to cope with long-term occlusion; for tracks with extremely low scores or exhibiting non-physical motion characteristics, they are identified as clutter and memory resources are released immediately to ensure that the single-frame association time is maintained at the millisecond level.

10. The low-altitude radar multi-target secondary tracking method based on trajectory-level correlation and robust features as described in claim 9, characterized in that, Step S1 involves processing the complex-valued signal output by the radar to obtain slant range, azimuth, pitch, and Doppler velocity, which are then input into a primary tracker for real-time state estimation, and the resulting track fragment with initial tags is output.

Images