Radar signal multi-target tracking method and system based on dynamic graph convolutional network

Abstract

The invention provides a radar signal multi-target tracking method and system based on a dynamic graph convolutional network. The method comprises the following steps: acquiring a radar one-dimensional range profile sequence of a target cluster, and performing pulse compression to form a radar echo two-dimensional sequence; separating position and amplitude data of aliasing scattering points based on scattering point distribution characteristics; calculating a spatial distance between targets by using position information, constructing a dynamic edge weight in combination with a radial speed difference, and generating a shielding relation edge; aggregating neighborhood features through a dynamic graph convolutional network, and learning trajectory decoupling features of the occlusion and intersection regions; and generating association weights based on the decoupling features, inputting the association weights into a multi-hypothesis tracking algorithm, generating a multi-branch trajectory, and if the spatial distance between two branches is smaller than a dynamic threshold value and the velocity vector included angle exceeds a set angle, retaining a branch with a higher association weight, and outputting a multi-target trajectory sequence. According to the invention, the multi-target tracking precision and robustness in a shielding and trajectory crossing scene are significantly improved.

Description

Technical Field

[0001] The present application relates to the field of radar signal processing technology, and in particular to a radar signal multi-target tracking method and system based on a dynamic graph convolutional network. Background Art

[0002] In high-density target scenarios like drone swarm tracking and vehicle formation monitoring, radar systems must address the problem of aliasing caused by closely spaced targets, leading to loss of target position and identity information. Trajectory intersections can cause track confusion, making it difficult for traditional algorithms to decouple overlapping features, resulting in a sharp drop in tracking accuracy. Both military and transportation applications require systems to maintain real-time responsiveness despite strong clutter interference, while ensuring trajectory continuity and identity consistency.

[0003] Currently, the mainstream approach utilizes a multi-hypothesis tracking (MHT) framework combined with an interactive multi-model filter (IMF). The MHT algorithm constructs track hypothesis branches based on distance and velocity correlation thresholds, selecting the most probable trajectory. The IMF runs multiple motion models in parallel, weighting the target's motion state based on model probabilities. The dynamically updated state estimate adapts to the maneuvering target's behavior.

[0004] However, this approach exhibits significant shortcomings in scenarios with high-density occlusion and intersecting trajectories. Occlusion-induced measurement loss leads to a surge in the number of associated hypotheses, resulting in heavy computational burdens, frequent erroneous branch selection, and significantly exacerbated trajectory confusion. Filters relying on pre-defined models struggle to respond to instantaneous target maneuvers, and delayed motion state updates lead to significant increases in position estimation errors. Summary of the Invention

[0005] The present application provides a radar signal multi-target tracking method and system based on a dynamic graph convolutional network to solve the problem of high trajectory breakage rate in the prior art.

[0006] In a first aspect, the present application provides a radar signal multi-target tracking method based on a dynamic graph convolutional network, comprising:

[0007] Collecting a one-dimensional radar range image sequence of the target cluster, performing pulse compression processing on the one-dimensional range image sequence, and arranging it in a slow time dimension to form a two-dimensional radar echo sequence;

[0008] Separating the positions and amplitudes of aliased scattering points in the two-dimensional radar echo sequence according to the linear or curved distribution characteristics of the scattering points of the cone target to obtain scattering point position and amplitude data;

[0009] Based on the position information in the scattering point position and amplitude data, the spatial distance between targets is calculated, and the dynamic edge weight is constructed in combination with the radial velocity difference. When the spatial distance between targets is less than the radar resolution threshold, an occlusion relationship edge is generated;

[0010] Initialize node feature weights according to amplitude information in the scatter point position and amplitude data, take the dynamic edge weight as a neighborhood aggregation coefficient of a graph attention mechanism, aggregate neighborhood node features of the occlusion relationship edge through a dynamic graph convolution network, learn decoupling features of a target trajectory in an occlusion and intersection area, and generate trajectory decoupling features.

[0011] Generate correlation weights based on the trajectory decoupling features, input a multi-hypothesis tracking algorithm to generate multi-branch trajectory hypotheses, perform conflict resolution based on an instantaneous speed vector angle of the multi-branch trajectory hypotheses, when a spatial distance of two branches is less than a dynamic threshold and the speed vector angle exceeds a set angle, retain a branch trajectory hypothesis with a higher correlation weight, and output a multi-target trajectory sequence.

[0012] Optionally, initializing node feature weights according to amplitude information in the scatter point position and amplitude data, taking the dynamic edge weight as a neighborhood aggregation coefficient of a graph attention mechanism, aggregating neighborhood node features of the occlusion relationship edge through a dynamic graph convolution network, learning decoupling features of a target trajectory in an occlusion and intersection area, and generating trajectory decoupling features, include:

[0013] Converting an amplitude value to an initial node feature vector based on amplitude information in the scatter point position and amplitude data, and initializing node feature weights based on the initial node feature vector;

[0014] Assigning the dynamic edge weight to a neighborhood aggregation coefficient of a graph attention mechanism, performing a linear transformation operation on the initial node feature vector through a learnable weight matrix, and generating an intermediate node feature;

[0015] Through a dynamic graph convolution network, taking the occlusion relationship edge as a connection path, aggregating neighborhood node features by weighted summing intermediate node features of neighbors of the occlusion relationship edge;

[0016] Concatenating the neighborhood node features after weighted summing and the intermediate node feature of the current node, performing a nonlinear activation operation on the concatenated features, learning decoupling features of a target trajectory in an occlusion and intersection area, and outputting an updated feature vector as a trajectory decoupling feature.

[0017] Optionally, calculating an inter-target spatial distance based on position information in the scatter point position and amplitude data, and constructing a dynamic edge weight in combination with a radial velocity difference, when the inter-target spatial distance is less than a radar resolution threshold, generating an occlusion relationship edge, including:

[0018] Calculating an inter-target spatial distance between any two scatter points based on position information in the scatter point position and amplitude data, and calculating a radial velocity difference value of a scatter point according to a change in the position information in a slow time dimension;

[0019] multiplying the inter-target spatial distance with the radial velocity difference, and constructing a dynamic edge weight based on a result of the multiplication;

[0020] when the inter-target spatial distance is less than a radar resolution threshold, generating an occlusion relationship edge between the corresponding scattering points, and taking the dynamic edge weight as a correlation strength value of the occlusion relationship edge.

[0021] Optionally, an association weight is generated based on the trajectory decoupling feature, a multi-hypothesis tracking algorithm is input to generate a multi-branch trajectory hypothesis, conflict resolution is performed based on an instantaneous velocity vector angle of the multi-branch trajectory hypothesis, when the spatial distance of two branches is less than a dynamic threshold and the instantaneous velocity vector angle exceeds a set angle, a branch trajectory hypothesis with a higher association weight is retained, and a multi-target trajectory sequence is output, including:

[0022] based on the trajectory decoupling feature, a feature vector similarity of different trajectory points is calculated, an association weight is generated, and the association weight is input to a multi-hypothesis tracking algorithm to generate a multi-branch trajectory hypothesis;

[0023] for the multi-branch trajectory hypothesis, when the spatial distance of any two branches is less than a dynamic threshold, an instantaneous velocity vector angle of the two branches is further calculated, when the instantaneous velocity vector angle exceeds a set angle, the association weights of the two branches are compared, a branch trajectory hypothesis with a higher association weight is retained, the retained branch trajectory hypothesis is integrated, and a multi-target trajectory sequence is output in time sequence.

[0024] Optionally, for the multi-branch trajectory hypothesis, when the spatial distance of any two branches is less than a dynamic threshold, an instantaneous velocity vector angle of the two branches is further calculated, when the instantaneous velocity vector angle exceeds a set angle, the association weights of the two branches are compared, a branch trajectory hypothesis with a higher association weight is retained, the retained branch trajectory hypothesis is integrated, and a multi-target trajectory sequence is output in time sequence, including:

[0025] the multi-branch trajectory hypothesis with a spatial distance less than a dynamic threshold at the current time is screened to form a set of branch pairs to be processed, and instantaneous velocity vectors of each branch trajectory pair in the set of branch pairs to be processed at the current time are extracted;

[0026] an angle value of the two instantaneous velocity vectors in the same branch trajectory pair is calculated, and the angle value is obtained by vector dot product operation and inverse sine function conversion;

[0027] when the angle value exceeds a set angle, the association weight values of the two branch trajectories are extracted from the set of branch pairs to be processed; the association weight values of the two branch trajectories are compared, the branch trajectory with a larger association weight value is retained, and the branch trajectory with a smaller association weight value is removed;

[0028] The branch trajectory hypothesis after integration is reserved, and a multi-target trajectory sequence is output in chronological order.

[0029] Optionally, according to the linear or curved surface distribution characteristics of the scattering points of the cone target, the positions and amplitudes of the mixed scattering points in the radar echo two-dimensional sequence are separated to obtain scattering point position and amplitude data, including:

[0030] The position information of each scattering point in the radar echo two-dimensional sequence in the slow time dimension is extracted, the continuous shift pattern of the change sequence is analyzed, and the position distribution characteristics of the change sequence are identified as linear distribution characteristics or curved surface distribution characteristics;

[0031] For the linear distribution characteristics of the scattering points of the cone target, according to the continuity of the adjacent slow time frame position shift direction and the smoothness of the amplitude change, the scattering points satisfying the continuity condition are divided into the same target;

[0032] For the curved surface distribution characteristics of the scattering points of the cone target, the curvature change value of the scattering point position change sequence is calculated, and when the curvature change value is less than a preset threshold and the position shift amplitude remains continuously increasing, it is determined that the scattering point belongs to the same target;

[0033] Based on the scattering points divided into the same target, the spatial position overlapping position information is detected, the mixed scattering point position and amplitude are separated, and the separated scattering point position and amplitude data are obtained.

[0034] Optionally, through a dynamic graph convolution network, the occlusion relationship edge is used as a connection path, the intermediate node features of the neighborhood of the occlusion relationship edge are summed by weighting, and the neighborhood node features are aggregated, including:

[0035] Based on the occlusion relationship edge, all neighborhood nodes directly connected to the current node through the occlusion relationship edge are screened out to form a neighborhood node set;

[0036] The intermediate node features of each node in the neighborhood node set are extracted, and the intermediate node features are generated by linear transformation from the initial node features;

[0037] The dynamic edge weight is used as a neighborhood aggregation coefficient, and the intermediate node features of each neighborhood node are multiplied by the corresponding neighborhood aggregation coefficient to generate weighted neighborhood node features;

[0038] The weighted neighborhood node features of the same current node are summed to output the aggregated neighborhood node features.

[0039] In a second aspect, the application provides a radar signal multi-target tracking system based on a dynamic graph convolution network, including:

[0040] An acquisition module is used to acquire a one-dimensional radar range image sequence of a target cluster, perform pulse compression processing on the one-dimensional range image sequence, and arrange it in a slow time dimension to form a two-dimensional radar echo sequence;

[0041] A separation module is used to separate the positions and amplitudes of aliased scattering points in the two-dimensional sequence of radar echoes according to the linear or curved distribution characteristics of the scattering points of the cone target, so as to obtain scattering point position and amplitude data;

[0042] A construction module is configured to calculate the spatial distance between targets based on the position information in the scattering point position and amplitude data, and to construct a dynamic edge weight based on the radial velocity difference, and to generate an occlusion relationship edge when the spatial distance between targets is less than a radar resolution threshold;

[0043] A generation module is configured to initialize node feature weights based on the scattering point positions and amplitude information in the amplitude data, use the dynamic edge weights as neighborhood aggregation coefficients of a graph attention mechanism, aggregate neighborhood node features of the occlusion relationship edges through a dynamic graph convolutional network, learn the decoupling features of the target trajectory in the occlusion and intersection areas, and generate trajectory decoupling features;

[0044] The output module is used to generate association weights based on the trajectory decoupling features, input the multi-hypothesis tracking algorithm to generate multi-branch trajectory hypotheses, and perform conflict resolution based on the instantaneous velocity vector angles of the multi-branch trajectory hypotheses. When the spatial distance between the two branches is less than a dynamic threshold and the velocity vector angle exceeds a set angle, the branch trajectory hypothesis with a higher association weight is retained, and a multi-target trajectory sequence is output.

[0045] In a third aspect, the present application provides a computing device comprising a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are used to be called and executed by the processing component to implement a radar signal multi-target tracking method based on a dynamic graph convolutional network as described in the first aspect above.

[0046] In a fourth aspect, the present application provides a computer storage medium storing a computer program. When the computer program is executed by a computer, it implements a radar signal multi-target tracking method based on a dynamic graph convolutional network as described in the first aspect.

[0047] This application collects a one-dimensional radar range image sequence of a target cluster and performs pulse compression processing to form a two-dimensional radar echo sequence. This step significantly improves the signal-to-noise ratio and resolution, laying the foundation for subsequent operations; then, the position and amplitude data of the aliased scattering points are separated according to the scattering point distribution characteristics of the conical target, effectively avoiding target position ambiguity and accurately extracting parameters; then, the spatial distance between targets is calculated based on the position information, and the dynamic edge weight is constructed in combination with the radial velocity difference. When the distance is less than the radar resolution, the occlusion relationship edge is generated, accurately quantifying the relationship between targets and reducing trajectory confusion; then, the node feature weight is initialized according to the amplitude information, and the dynamic edge weight is used as the aggregation coefficient of the graph attention mechanism. The decoupled features are learned by aggregating the neighborhood node features through the dynamic graph convolutional network, effectively decoupling the trajectory features of the occlusion and intersection areas and solving the fracture problem; finally, the correlation weight is generated based on the decoupled features and input into the multi-hypothesis tracking algorithm. After the trajectory is generated, the conflict is resolved based on the velocity vector angle. When the spatial distance is small and the angle is large, the high-weight branch is retained to output the trajectory sequence, significantly reducing the false detection rate and identity switching rate.

[0048] Furthermore, based on the amplitude information of the scattering points, the amplitude values ​​are converted into initial node feature vectors and the node feature weights are initialized, retaining the target physical properties to prevent feature loss; then, the dynamic edge weights are assigned to the neighborhood aggregation coefficient of the graph attention mechanism, and the features are linearly transformed through the learnable weight matrix to generate intermediate node features, thereby improving the adaptability to occlusion relationships; subsequently, through the dynamic graph convolutional network, the occlusion relationship edges are used as the connection paths, and the weighted summation of the neighborhood node features is performed to achieve aggregation, effectively compensating for feature transmission interruptions and maintaining trajectory continuity; finally, the aggregated features are spliced ​​with the intermediate node features, and a nonlinear activation operation is performed to learn the decoupled features, and the updated feature vector is output as the trajectory decoupling feature, which significantly enhances the trajectory recognition ability of occlusion and intersection areas and improves the tracking accuracy.

[0049] These and other aspects of the present application will become more readily apparent from the description of the following embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0051] Figure 1 A flowchart of a radar signal multi-target tracking method based on a dynamic graph convolutional network provided by the present application is shown;

[0052] Figure 2A scene diagram showing a radar signal multi-target tracking method based on a dynamic graph convolutional network provided by the present application is shown;

[0053] Figure 3 The present invention provides a schematic diagram of a radar signal multi-target tracking system based on a dynamic graph convolutional network.

[0054] Figure 4 A schematic structural diagram of a computing device provided by the present application is shown. DETAILED DESCRIPTION

[0055] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0056] In some of the processes described in the specification and claims of this application and the above-mentioned figures, multiple operations that appear in a specific order are included, but it should be clearly understood that these operations may not be executed in the order in which they appear in this document or may be executed in parallel. The serial numbers of the operations, such as 101, 102, etc., are only used to distinguish between different operations, and the serial numbers themselves do not represent any order of execution. In addition, these processes may include more or fewer operations, and these operations may be executed in sequence or in parallel. It should be noted that the descriptions of "first", "second", etc. in this document are used to distinguish different messages, devices, modules, etc., and do not represent a sequential order, nor do they limit "first" and "second" to being different types.

[0057] Researchers have found that in the field of radar multi-target tracking, existing methods have significant shortcomings in scenarios with high-density occlusion and intersecting trajectories: the measurement loss caused by occlusion leads to a surge in the number of associated hypotheses, a heavy computational burden, frequent incorrect branch selection, and exacerbated trajectory confusion. At the same time, motion filters that rely on preset models have difficulty responding to instantaneous target maneuvers, and delayed motion state updates lead to a significant increase in position estimation errors. For example, in a dense target cluster environment, traditional methods cannot effectively separate aliased scattering points, increasing the risk of trajectory mismatch and position deviation. Therefore, a multi-target tracking method for radar signals based on a dynamic graph convolutional network is urgently needed.

[0058] To address the above issues, the present invention proposes a radar signal multi-target tracking method based on a dynamic graph convolutional network. The core of the method is to dynamically model the occlusion relationship and intersection characteristics between targets using dynamic edge weights and an attention mechanism to achieve decoupled learning and efficient updating of trajectory features. Specifically, a one-dimensional radar range image sequence of a target cluster is collected and pulse compressed to form a two-dimensional sequence. The position and amplitude of aliased scattering points are separated according to the scattering point distribution characteristics of the cone target. Dynamic edge weights are calculated based on the position information to construct the spatial distance and radial velocity difference between targets. When the distance is less than the radar resolution threshold, an occlusion relationship edge is generated. Node feature weights are initialized, and the dynamic edge weights are used as the neighborhood aggregation coefficient of the graph attention mechanism. The decoupled features of the target trajectory in the occlusion and intersection areas are learned by aggregating the neighborhood node features of the occlusion edge through a dynamic graph convolutional network. Based on the decoupled features, association weights are generated and input into a multi-hypothesis tracking algorithm to generate multiple branch trajectory hypotheses. The instantaneous velocity vector angle conflict is resolved, and the branch trajectory with higher weight is retained to output the trajectory sequence. This method effectively solves the problem of a surge in associated hypotheses in scenarios with high-density occlusion and trajectory intersections, reduces the computational burden and the frequency of erroneous branch selection, alleviates trajectory confusion, and improves motion state response speed through dynamic modeling and immediate conflict resolution, avoids position estimation bias, and significantly enhances the robustness and accuracy of multi-target tracking.

[0059] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without making creative efforts are within the scope of protection of this application.

[0060] Figure 1 The present invention provides a flowchart of a radar signal multi-target tracking method based on a dynamic graph convolutional network. Figure 1 As shown, the method includes:

[0061] 101. Collect a one-dimensional radar range image sequence of the target cluster, perform pulse compression processing on the one-dimensional range image sequence, and arrange them in a slow time dimension to form a two-dimensional radar echo sequence;

[0062] In the above steps, a target cluster refers to a collection of multiple target objects detected by the radar, such as a group of aircraft or vehicles. The radar one-dimensional range profile sequence is a series of one-dimensional range data collected by the radar system in the slow time dimension. Each data point corresponds to a target range information sequence received after a pulse is transmitted. This sequence is arranged in slow time order and contains the range distribution and amplitude of the target. Pulse compression processing is the process of filtering the radar echo signal using matched filter technology. By compressing the pulse width, the range resolution is improved, thereby distinguishing close-range targets. The slow time dimension represents the time scale of the radar pulse repetition interval (PRI), which corresponds to the time interval change between different pulse transmissions. The radar echo two-dimensional sequence is a two-dimensional data structure formed after pulse compression processing. It is arranged in a matrix form with slow time as the row index and distance as the column index for subsequent signal analysis.

[0063] In an embodiment of the present application, a radar one-dimensional range image sequence of a target cluster is first collected, and pulse compression processing is performed to form a two-dimensional sequence of radar echoes; for example, in a radar system, a sequence of radar pulse signals such as a linear frequency modulation signal is transmitted to the target cluster area. After each pulse is transmitted, the receiving system captures the echo signal and converts it into a digital data sequence. Each sequence is called a one-dimensional range image. Assume that the system transmits 20 pulses, and the echo sequence corresponding to each pulse contains 256 sampling points representing the values ​​of different distance units. After collection, a sequence of 20 256-point range images is formed; then pulse compression processing is applied to each one-dimensional range image, and a matched filter algorithm is used to improve the range resolution. The specific process is to convert the received signal into the frequency domain: first, a fast Fourier transform (FFT) is performed on the input signal to obtain a spectrum, and then it is multiplied with a pre-generated reference signal such as the complex conjugate spectrum of the transmitted pulse. Here, the reference signal is obtained by sampling and numerical calculation. The matched filter output is expressed as Y(k) = X(k)·H*(k), where X(k) is the FFT spectrum of the received signal, H*(k) is the complex conjugate spectrum of the reference signal, and k is the frequency domain index. Finally, Y(k) is converted back to the time domain signal through an inverse fast Fourier transform (IFFT), thereby obtaining a compressed range image. For example, the echo sequence data of the first pulse, such as [1.2, 0.8, 0.5], is assumed to be in units of amplitude values, and the reference signal spectrum H*(k) is assumed to be [0.9, 0.7, 0.6]. After FFT calculation, the spectra are multiplied and then IFFT is performed to return a compressed sequence such as [1.1, 0.85, 0.55], significantly improving the resolution. Finally, all processed range images are rearranged according to the slow time dimension, that is, the pulse transmission order. For example, 20 compressed 256-point sequences are combined into a 20-row, 256-column two-dimensional matrix, which serves as the radar echo two-dimensional sequence for subsequent signal processing steps.

[0064] In a practical application, a radar detection experiment involved a scenario involving occlusion and intersecting trajectories of a cluster of targets in mid-air. Radar signal processing methods were employed to address this challenge. In the experiment, a radar was located at observation point A, pointing toward the sky above area B, where multiple targets, such as Target C and Target D, were present. These targets obscured each other during flight and their trajectories frequently intersected, making it difficult to discern target details using a single radar measurement. First, the radar continuously transmitted and received echo signals to acquire a one-dimensional range profile sequence of the target cluster. Specifically, since the radar transmit pulse repetition frequency was 10 kHz and the acquisition duration was 10 seconds, the total number of pulses was calculated to be 100 pulses (repetition frequency × acquisition time = 10 kHz × 10). This yielded a one-dimensional range profile sequence corresponding to each pulse, containing signal distribution data in the range dimension. Each one-dimensional range profile sequence is then pulse compressed to improve the target's resolution in the range dimension. During this processing, each sequence is input with a length of 200 sampling points, and a preset matched filter is used for calculation. For example, for each point in the sequence, the filter output value = the input signal point × the reference signal point integral value. This point-by-point operation generates a compressed, high-resolution range profile. Finally, these processed sequences are arranged along the slow-time dimension, or time axis, to form a two-dimensional sequence of radar echoes. The vertical dimension represents the time step, with a total of 100 time points corresponding to each pulse acquisition moment, and the horizontal dimension represents the range sampling points, totaling 200 points. The resulting matrix forms a 100×200 two-dimensional matrix, facilitating subsequent analysis of target dynamic behavior under occlusion and intersecting tracks.

[0065] In the overall solution of step 101 above, by collecting a one-dimensional radar range image sequence of the target cluster, pulse compression technology is used to significantly improve signal resolution and signal-to-noise ratio, effectively suppress interference sidelobes and enhance target feature extraction capabilities; then, a structured two-dimensional sequence of radar echoes is constructed by precise arrangement according to the slow time dimension. This sequence completely retains the coupled motion characteristics of the target in the radial range and slow time domains, providing a phase-consistent data base for high-precision cluster target imaging, while greatly improving subsequent motion trajectory analysis capabilities.

[0066] 102. Separate the positions and amplitudes of aliased scattering points in the two-dimensional radar echo sequence according to the linear or curved distribution characteristics of the scattering points of the cone target to obtain scattering point position and amplitude data;

[0067] Optionally, step 102 may specifically include the following steps:

[0068] 1021. Extract a change sequence of the position information of each scattering point in the two-dimensional radar echo sequence in a slow time dimension, analyze a continuous offset pattern of the change sequence, and identify whether the position distribution characteristic of the change sequence is a linear distribution characteristic or a curved surface distribution characteristic;

[0069] 1022. Based on the linear distribution characteristics of the scattering points of the cone target, the scattering points that meet the continuity condition are classified into the same target according to the continuity of the position offset directions of adjacent slow time frames and the smoothness of the amplitude change;

[0070] 1023. Calculate a curvature change value of a position change sequence of the scattering points based on the surface distribution characteristics of the scattering points of the cone target. When the curvature change value is less than a preset threshold and the position offset amplitude continues to increase, determine that the scattering points belong to the same target.

[0071] 1024. Based on the scattering points classified into the same target, detect the position information of the overlapping spatial positions, separate the aliased scattering point positions and amplitudes, and obtain separated scattering point position and amplitude data.

[0072] In the above steps, cone target refers to a conical structure object detected by radar, such as a missile warhead or similar device; scattering point refers to a discrete position point on the target surface that reflects radar waves, and each point contains position and amplitude information; linear distribution characteristics describe the linear change trend of the scattering point position in the slow time dimension, that is, at different pulse emission time points; surface distribution characteristics describe the curved path of the scattering point position change; aliased scattering point refers to a scattering point where radar echoes overlap and are difficult to distinguish at the same position due to the close distance between multiple targets; slow time dimension is a time scale with radar pulse repetition interval as the unit, which is used to arrange pulse sequences; position information is the coordinate data of the scattering point in different slow time frames, such as distance value; amplitude data represents the scattering point The intensity value of the reflected signal; slow time frame refers to the complete data set corresponding to a single pulse echo, including multiple scattering points; continuity of position offset direction refers to the fact that the movement direction of the scattering points between adjacent slow time frames, such as positive or negative, remains consistent; smoothness of amplitude change refers to the fact that the amplitude values ​​of the scattering points in adjacent frames fluctuate little and the changes are continuous; curvature change value is a numerical measure that quantifies the degree of curvature of the position change path, in radians, obtained through mathematical calculation; preset threshold is a critical value set for judging curvature stability or amplitude change, such as a fixed number; spatial position overlap refers to the fact that scattering points of different targets appear at the same distance unit position; target is the individual object entity being detected; these concepts form the basis for subsequent processing of scattering point data.

[0073] In the embodiment of the present application, first, step 1021 is used to extract the position change sequence of each scattering point in the two-dimensional sequence of radar echoes and analyze the distribution characteristics; the position change sequence of each scattering point is a continuous position value extracted from the slow time dimension, for example, in a slow time sequence, the position of the scattering point gradually moves from 150 meters in the first frame to 152 meters in the second frame, and then to 154 meters in the third frame; then, the continuous offset pattern of the change sequence is analyzed, and the characteristics are determined by fitting the position points by the least squares method. If the fitted straight line is such as the function y=at+b, where a and b are the residual values ​​of the fitting parameters, and the sum of the squares of the differences between the actual position and the fitted value is less than a preset threshold, such as 5% of the total distance , it is identified as a linear distribution characteristic, otherwise a quadratic function is used to fit the curve. If the residual is stable, it is a surface distribution characteristic; for example, a scattering point position sequence is 150, 152, and 154 meters in three slow time frames. A linear fit is used to obtain the function y=2t+148 (t is the time index). The residual is calculated. For example, the actual value of point 150 meters and the fitted value 148 have a square difference of 2, which is 4, the square difference of point 152 is 0, and the square difference of point 154 is 0. The total residual value is 4. The threshold is set to 5% of the difference of 4 meters in the position value range from 150 to 154, that is, 0.2 meters. The residual 4 is greater than the threshold 0.2, so the quadratic function is used to fit the curve equation such as y=0.05t 2 +1.9t+149, the new residual 0.1 meter is less than the threshold 0.2 meter, and it is determined to be a surface distribution feature.

[0074] Secondly, in step 1022, the scattering points with linear distribution characteristics are divided into the same target; based on the continuity detection of the position offset direction, the position difference of adjacent slow time frames is calculated, such as the offset from the first frame to the second frame is 152-150=2 meters (positive direction), and the offset from the second frame to the third frame is 154-152=2 meters (positive direction). If the offset directions of three or more consecutive frames are consistent, all positive or all negative, for example, the positive movement sequence remains stable; at the same time, the smoothness of the amplitude change is checked, and the absolute percentage of the amplitude difference between adjacent frames is calculated, such as the amplitude of the first frame is 0.8, the second frame is 0.82, and the third frame is 0.8. 3 frames are 0.81, the change rate from the 1st to the 2nd frame is |0.82-0.8| / 0.8=2.5%, and from the 2nd to the 3rd frame is |0.81-0.82| / 0.82≈1.22%. If the change rate is less than a threshold value such as 5% and three consecutive frames meet the conditions, then these scatter points are judged to belong to the same target. For example, the sequence position change of target A is [150,152,154] meters (continuous positive) and the amplitude is [0.8,0.82,0.81] (fluctuation is less than 5%), while the sequence of target B is [140,138,136] meters (negative movement) and they are considered different targets.

[0075] Next, in step 1023, the scattering points with surface distribution characteristics are divided into the same target; the curvature change value of the scattering point position change sequence is calculated, and the coordinates of three consecutive points are taken from the position sequence, such as the first frame P1 (150 meters, amplitude 0.8), the second frame P2 (152 meters, amplitude 0.82), and the third frame P3 (155 meters, amplitude 0.83), where the position coordinates are regarded as two-dimensional point vectors; then the vector cross product modulus divided by the modulus length is calculated as the curvature, and the formula is: Among them, Δr=P2-P1 (coordinate difference vector, such as (2 meters, 0.02 amplitude)), Δr'=P3-P2 coordinate difference vector, such as (3 meters, 0.01 amplitude), cross product modulus is the result of cross operation of vector components, such as |2×0.01-0.02×3|=|0.02-0.06|=0.04 unit meter·amplitude, and the cube of modulus length is the cube of the length of each vector; the specific process is: first calculate the vector difference, then calculate the cross product modulus, and finally divide by the cube of the modulus length to get the curvature value; when five consecutive If the curvature values ​​of the frames are all less than a preset threshold, such as 0.2 radians, and the position offset amplitude keeps increasing continuously (for example, the position value increases monotonically), then the scattering points are determined to belong to the same target. For example, taking three frames of data P1 (150, 0.8), P2 (152, 0.82), and P3 (155, 0.83), the vectors Δr = (2, 0.02) and Δr' = (3, 0.01) are calculated, and the cross product modulus = |2·0.01-0.02·3| = |0.02-0.06| = 0.04, Unit: meter, modulus length cubic (2.0002) 3 ≈8.0024, curvature The arc angle is less than the threshold of 0.2, and the position increases continuously from 150 to 155, so it is classified as the same target, and similar sequences are repeatedly checked to confirm that five consecutive frames meet the conditions.

[0076] Finally, step 1024 is used to separate aliased scattering points based on the scattering points classified as the same target and obtain position and amplitude data; points with overlapping spatial positions are detected within the group, for example, scattering points of different targets are located in the same distance unit, such as 150 meters, in the same slow time frame; aliased points are separated by analyzing amplitude changes, and when the amplitude of the overlapping point suddenly changes beyond the smoothness threshold, such as when the adjacent frames change by more than 10%, the true position is reconstructed using interpolation; specifically, at the same distance unit, if the amplitude values ​​of adjacent frames suddenly increase, such as 0.8 at 150 meters in the first frame and 1.6 in the second frame, it means that two scattering points are aliased, and the cubic spline interpolation algorithm is used to reconstruct the true position from the surrounding The position data is reconstructed into a curve, and the position offset is decomposed into two points, 150.2 meters and 149.8 meters, with the amplitudes divided according to the original ratio. For example, in the target group, the aliased point with an initial amplitude of 1.6 is located at 150 meters. Assuming that the position and amplitude trends of the surrounding points show a reasonable split ratio of 0.5:0.5, two scattering points are obtained after separation, with positions of 149.8 meters (amplitude 0.8) and 150.2 meters (amplitude 0.8). The separated scattering point positions and amplitude matrix data are finally output. For example, the output position sequence of target 1 is [149.8, 151.9, ...] meters and the amplitude sequence is [0.8, 0.82, ...].

[0077] In practical applications, in certain aerial target detection scenarios involving mutual occlusion and trajectory intersection among spatial clusters of targets in flight, a radar system uses an existing two-dimensional echo sequence, such as the 100×200 matrix output from step 101, with the vertical dimension representing 100 time steps and the horizontal dimension representing 200 range sampling points, to address the issue of aliased scattering points and isolate the characteristic information of different targets. First, the temporal variation sequence of the position information of each scattering point in the two-dimensional sequence is extracted. For each scattering point in the sequence, its position offset values ​​at all time steps are selected to form a sequence. For example, the position offset sequence of a typical scattering point might gradually change from a position of 10 units to a position of 50 units over 100 time steps. Next, the distribution characteristics of this variation sequence are analyzed. If the offset values ​​show a constant linear trend, such as a uniform increase or decrease in position, the distribution is considered linear. Otherwise, if the offset values ​​show a curvature or arc-shaped variation, the distribution is considered curved. For linear distribution characteristics, the continuity of the offset direction of adjacent time steps in the sequence is checked. For example, the offset direction from time step 10 to step 11 is consistently positive, and the amplitude change is smooth, such as the amplitude value smoothly changes from 1.0 units to 1.2 units. If these conditions are continuously met, the scatter points are classified as belonging to the same target group. For surface distribution characteristics, the curvature change value of the position change sequence is calculated. For example, the curvature change value = |curvature at time step 50 - curvature at time step 25|, with a preset threshold of 0.1 units. If the calculated curvature change value is less than 0.1 and the position offset amplitude continuously increases, such as the displacement value increases by an average of 0.5 units per step from time steps 1 to 100, with a total displacement of 50 units, the scatter points are judged to belong to the same target. Finally, based on the above division results, the position overlap information at a specific time step, such as time step 75, is detected. For example, two scattering points overlap near the position of 35 units. By adjusting the position and amplitude model, the aliased scattering points are separated, thereby obtaining clear and separated position and amplitude data, which facilitates subsequent target tracking.

[0078] In the overall solution of the above step 102, based on the unique scattering point distribution characteristics of the conical target, the radar echo two-dimensional sequence is refined and analyzed to first extract the scattering point position change sequence in the slow time dimension, and accurately identify its linear or curved surface distribution pattern; for the linear distribution characteristics, the continuity of the position offset of adjacent frames and the smooth change law of the amplitude are used to realize the attribution of the scattering points, and for the curved surface distribution characteristics, the curvature change threshold and the position offset continuity constraint are used to confirm the target association; finally, by detecting the spatial position aliasing points and combining the amplitude and position evolution characteristics, the scattering points are accurately separated, thereby obtaining scattering point position and amplitude data with clear physical meaning, spatiotemporal consistency and de-aliasing, providing a clear structural underlying feature support for subsequent high-precision three-dimensional reconstruction and motion inversion.

[0079] 103. Calculate the inter-target spatial distance based on the position information in the scatter point position and amplitude data, and construct a dynamic edge weight combined with the radial velocity difference, and generate a shielding relationship edge when the inter-target spatial distance is less than the radar resolution threshold;

[0080] Optionally, step 103 can specifically include the following steps:

[0081] 1031. Calculate the inter-target spatial distance between any two scatter points based on the position information in the scatter point position and amplitude data, and calculate the radial velocity difference value of the scatter point according to the change of the position information in the slow time dimension;

[0082] 1032. Perform a multiplication operation on the inter-target spatial distance and the radial velocity difference value, and construct a dynamic edge weight based on the multiplication operation result;

[0083] 1033. When the inter-target spatial distance is less than the radar resolution threshold, generate a shielding relationship edge between the corresponding scatter points, and the dynamic edge weight as the correlation strength value of the shielding relationship edge.

[0084] In the above steps, the inter-target spatial distance represents the actual spatial interval between any two scatter points, which is calculated based on the position information of the scatter points, for example, using the distance formula to calculate the unit meter; the position information is the coordinate data stored by the scatter point, including the position value at different slow time points; the slow time dimension is a time scale in units of radar pulse repetition interval, such as the time index of the pulse sequence; the radial velocity difference value is the numerical difference of the velocities of two scatter points in the radar observation direction, which is calculated by calculating the velocity change rate from the change sequence of the position information in the slow time dimension; the dynamic edge weight is the weight value of the constructed target relationship graph edge, which is obtained by mathematical operation and is used to quantify the relationship strength; the multiplication operation is an algorithm for multiplying two numerical values, which specifically adopts multiplication calculation; the radar resolution threshold is the minimum distance value that the radar system can distinguish between targets, for example, a pre-set critical value such as 0.5 meters; the shielding relationship edge is a connecting line in the radar target relationship graph representing the shielding between targets; the correlation strength value is a numerical weight stored in the shielding relationship edge, reflecting the importance degree of the shielding relationship; these concepts are used for subsequent processing of the interaction between targets.

[0085] In the embodiments of the present application, firstly, the inter-target spatial distance between any two scattering points is calculated based on the position information in the scattering point position and amplitude data, and the radial velocity difference of the scattering points is calculated according to the change of the position information in the slow time dimension through step 1031. The specific process is as follows: the position sequence is extracted from the separated scattering point data, for example, the position sequence of a scattering point of target A is 100 meters in the first frame, 102 meters in the second frame and 104 meters in the third frame, and the position sequence of a scattering point of target B is 110 meters in the first frame, 112 meters in the second frame and 114 meters in the third frame; when calculating the spatial distance, the data of the same slow time frame is taken, and the Euclidean distance formula is used for calculation, the formula is d = |x1-x2|, wherein x1 and x2 are the position values (unit: meter) of the two scattering points at the same time point, for example, the position difference of the scattering points of target A and target B in the first frame is |100-110| = 10 meters; when calculating the radial velocity difference, the position change rate of the velocity of each scattering point is first calculated, for example, the velocity of the scattering point of target A is calculated as the position sequence difference divided by the time interval, assuming that the pulse interval is a fixed value of 0.1 second, the formula is wherein Δx is the position change value, Δt is the slow time frame interval time (unit: second), the position difference of target A from the first frame to the second frame is 102-100 = 2 meters, the time interval is 0.1 second, the velocity is meters / second, and the velocity of the scattering point of target B is calculated in the same way meters / second; then the radial velocity difference is calculated, the formula is Δv = |v A -v B |, for example, the difference |20-20| = 0 meters / second; finally, the distance and velocity difference matrix of all scattering point pairs is output, for example, the pairing result of target A-B is: spatial distance 10 meters, radial velocity difference 0 meters / second.

[0086] Secondly, the inter-target spatial distance and the radial velocity difference are multiplied through step 1032, and the dynamic edge weight is constructed based on the multiplication result; the specific process is as follows: from the data output from step 1031, the calculated values of the spatial distance and the radial velocity difference of each scattering point pair are taken to perform direct multiplication, the formula is w = d x |Δv|, wherein w is the dynamic edge weight value (unit: meter 2 / second), d is the spatial distance (unit: meter), |Δv| is the absolute value of the velocity difference (unit: meter / second), and the product value is used to quantify the strength of the dynamic relationship between targets; for example, from the pairing of target A-B in the previous example: distance 10 meters, velocity difference 0 meters / second, w = 10 x |0| = 0 meter 2 / second, indicating that the weight value is 0; for another pairing of target C-D, the distance is 5 meters and the velocity difference is 2 meters / second, w = 5 x 2 = 10 meter 2 / second, the output weight value is 10; when constructing, this multiplication is repeated for all scattering point pairs, and the weight value is stored as a dynamic edge weight matrix for subsequent steps.

[0087] Finally, by step 1033, when the target inter-space distance is less than the radar resolution threshold, an occlusion relationship edge is generated between the corresponding scattering points, and the dynamic edge weight is taken as the association strength value of the occlusion relationship edge; the specific process is: compare the spatial distance of each scattering point pair with the preset radar resolution threshold, for example, the system is set to 0.5 meters, if the distance is less than 0.5 meters, an occlusion relationship edge is generated to represent that the two targets will occlude each other; at the same time, the weight value of the pair is taken from the dynamic edge weight and directly assigned to the edge as the association strength value, reflecting the dynamic influence degree of the occlusion relationship; for example, the spatial distance of target A-B pair is 10 meters greater than the threshold 0.5 meters, no edge is generated; while the distance of target E-F pair is 0.3 meters less than the threshold 0.5 meters, a connection line from target E to target F is generated to represent the occlusion relationship, and the previously calculated dynamic edge weight value of the pair is taken, for example, the weight w = 8 meters is taken from the output of step 1032 2 / second, which is assigned as the association strength value; finally, the relationship graph structure is output, in which the number and strength of the edges are automatically constructed based on the above rules, for example, in a scene with 10 scattering points, if 5 pairs satisfy the distance less than the threshold, 5 edges are added with the association strength data.

[0088] In practical applications, in a certain radar monitoring scene, for the occlusion and trajectory intersection problem of the flying target group in the air, based on the separated scattering point data containing the position and amplitude information of 100 target scattering points, covering 100 time steps, the dynamic analysis of the inter-target relationship is carried out. For example, at time step 50, the scattering point position of target A is (30, 40, 50) units, and the scattering point position of target B is (32, 38, 48) units. First, the spatial distance of the two scattering points is calculated using the formula: Substituting the specific value, the calculation is units. At the same time, the radial velocity difference is calculated based on the change of position in the time dimension: target A changes from the position (20, 30, 40) units at time step 30 to the position (30, 40, 50) units at time step 50, the change time is 20 steps with an interval of 0.1 second, a total of 2 seconds, and the radial velocity is calculated as units / second; similarly, the radial velocity of target B is 7.5 units / second, so the velocity difference is |8.66-7.5| = 1.16 units / second. Then the dynamic edge weight is constructed: multiply the spatial distance 3.46 units by the velocity difference 1.16 units / second, the calculation process is weight = 3.46 x 1.16 ≈ 4.01 units 2Finally, we determine the blocking relationship and set the radar resolution threshold to 0.5 units. During the time step 40 to 60, we observe that the spatial distance between the two scattering points is continuously less than the threshold, for example, the distance value is in the range of 0.1 to 0.3 units. Therefore, we generate an blocking relationship edge between the scattering points and assign a dynamic edge weight of 4.01 units. 2 / second, as the correlation strength value, to describe the dynamic impact of targets in the process of mutual occlusion.

[0089] In the overall solution of step 103 above, by accurately analyzing the scattering point position and amplitude data, the spatial distance between targets is first calculated and the radial velocity difference is simultaneously extracted. The distance and velocity difference are innovatively multiplied to construct dynamic edge weights, and the weight values ​​can respond to changes in the target motion state in real time. When the target spacing exceeds the radar resolution threshold, an occlusion relationship edge with quantified correlation strength is automatically generated. This strength value integrates the target spatial position constraint and relative motion characteristics, effectively solving the problem of misjudgment of moving targets in traditional distance judgment, and forming a multi-dimensional fusion feature including target interaction strength and dynamic level of occlusion risk, providing key data support with physical significance and time adaptability for subsequent cluster target topology construction and occlusion compensation algorithm.

[0090] 104. Initialize node feature weights based on the scattering point positions and amplitude information in the amplitude data, use the dynamic edge weights as neighborhood aggregation coefficients of a graph attention mechanism, aggregate neighborhood node features of the occlusion relationship edges through a dynamic graph convolutional network, learn decoupling features of the target trajectory in the occlusion and intersection regions, and generate trajectory decoupling features.

[0091] Optionally, step 104 may specifically include the following steps:

[0092] 1041. Convert the amplitude value into an initial node feature vector based on the scattering point position and amplitude information in the amplitude data, and initialize the node feature weight based on the initial node feature vector;

[0093] 1042. Assign the dynamic edge weight to the neighborhood aggregation coefficient of the graph attention mechanism, perform a linear transformation operation on the initial node feature vector through a learnable weight matrix, and generate intermediate node features;

[0094] 1043. Using a dynamic graph convolutional network, taking the occlusion relationship edge as a connection path, performing weighted summation of intermediate node features in the neighborhood of the occlusion relationship edge, and aggregating neighborhood node features;

[0095] Among them, step 1043 may specifically include the following processes: based on the occlusion relationship edge, all neighboring nodes directly connected to the current node through the occlusion relationship edge are screened out to form a neighborhood node set; the intermediate node features of each node in the neighborhood node set are extracted, and the intermediate node features are generated by linear transformation of the initial node features; the dynamic edge weight is used as the neighborhood aggregation coefficient, and the intermediate node features of each neighborhood node are multiplied by the corresponding neighborhood aggregation coefficient to generate weighted neighborhood node features; all the weighted neighborhood node features of the same current node are summed up to output the aggregated neighborhood node features.

[0096] 1044. Concatenate the weighted summed neighborhood node features with the intermediate node features of the current node, perform a nonlinear activation operation on the concatenated features, learn the decoupling features of the target trajectory in the occlusion and intersection areas, and output the updated feature vector as the trajectory decoupling feature.

[0097] In the above steps, amplitude information refers to the echo signal strength value recorded in the scattering point data obtained in step 102; the initial node feature vector is a multi-dimensional numerical array converted by linear mapping of the amplitude value, which is used to represent the initial feature state of the target; the node feature weight is a numerical parameter assigned to each scattering point, indicating its importance in the network; the graph attention mechanism is an artificial intelligence model that can automatically focus on important nodes; the neighborhood aggregation coefficient is the weight ratio value used when aggregating neighbor node information; the dynamic edge weight is the relationship strength value output from step 103; the dynamic graph convolutional network is a neural network model that can handle dynamic connection relationships; the learnable weight matrix is ​​a transformation matrix automatically adjusted through training; the linear transformation operation is an algorithm that multiplies the input vector by the matrix; the intermediate node feature is the node feature generated after the linear transformation vector; neighborhood node features refer to the feature vectors of nearby nodes connected by occlusion relationship edges; weighted neighborhood node features are the results of neighborhood node features multiplied by coefficients; target trajectory refers to the target's motion path in time; occlusion and intersection areas describe the spatial positions of mutual influence between targets; decoupled features are independent target features separated after model learning; trajectory decoupled features are the final output feature vector representing the target path; connection path refers to the connection relationship between nodes formed by occlusion relationship edges; neighborhood node set is the set of all nodes directly connected to the current node; weighted summation is the process of multiplying multiple features by weights and then adding them up; splicing is the operation of connecting multiple feature sequences into a new sequence; nonlinear activation operation is the step of applying a nonlinear function to the input value; the updated feature vector is the new feature sequence output after network processing.

[0098] In the embodiment of the present application, first, step 1041 is used to convert the amplitude value into an initial node feature vector based on the scattering point position and the amplitude information in the amplitude data, and initialize the node feature weight; the specific process is: extract the amplitude value from the separated scattering point amplitude data, such as the amplitude of a scattering point of target A is 0.8, convert it into a vector of fixed dimension, assuming the dimension is 2, and the conversion formula is direct mapping, such as, the amplitude value a is converted into the vector [a, 0] to obtain the initial node feature vector, for example, the feature vector of target A is [0.8, 0]; then initialize the node feature weight, assign a preset initial value, such as the weight of all nodes is uniformly 0.5, indicating the basic importance in feature learning; for example, the scattering point amplitudes of targets A and B are 0.8 and 1.2 respectively, and the initial feature vectors after conversion are [0.8, 0] and [1.2, 0] respectively, and the weights of all nodes are initialized to 0.5.

[0099] Secondly, in step 1042, the dynamic edge weight is assigned to the neighborhood aggregation coefficient of the graph attention mechanism, and a linear transformation operation is performed on the initial node feature vector through the learnable weight matrix to generate the intermediate node feature; the specific process is: first, a value is taken from the dynamic edge weight matrix output in step 103, such as the weight of the target A and B pairing is 0.5, and it is stored as the neighborhood aggregation coefficient; then a learnable weight matrix is ​​applied, such as a randomly initialized matrix W, the size of which is assumed to be 2×2, and the linear transformation operation formula F is used. mid =W·F initial , where F initial is the initial node feature vector, W is the weight matrix, F mid is the intermediate node feature; for example, the initial feature of target A is [0.8,0], assuming the weight matrix W = [[1.0,0.5], [0.2,1.0]], calculate F mid =[1.0*0.8+0.5*0,0.2*0.8+1.0*0]=[0.8,0.16], the output is the intermediate node feature vector [0.8,0.16].

[0100] Next, in step 1043, the occlusion relationship edge is used as the connection path, and the neighborhood node features are aggregated through the dynamic graph convolutional network; the specific process is: first, based on the occlusion relationship edge generated in step 103, such as there is an edge between targets A and B, the neighboring nodes directly connected to the current node are screened to form a neighborhood node set, such as the current node is A and the neighbor is B; the intermediate node features of each node in the set are extracted, such as the feature [1.44, 0.24] of B from the output of step 1042; then the dynamic edge weight is used as the neighborhood aggregation coefficient, such as the AB edge weight 0.5, multiply each neighborhood node feature by the coefficient to generate a weighted neighborhood node feature, such as B's feature [1.44, 0.24] multiplied by 0.5 becomes [0.72, 0.12]; finally, sum the weighted neighborhood node features. If B has only one neighbor, directly output [0.72, 0.12] as the aggregated neighborhood node feature; for example, in a graph containing nodes A, B, and C, the neighborhood set of the current node A is B, and the weight of the AB edge is 0.5. After aggregation, the feature directly takes the weighted result [0.72, 0.12].

[0101] Finally, in step 1044, the weighted summed neighborhood node features are concatenated with the intermediate node features of the current node, and a nonlinear activation operation is performed on the concatenated features to learn the decoupling features of the target trajectory in the occlusion and intersection areas, and output the trajectory decoupling features. The specific process is: first, the intermediate node features of the current node, such as the feature [0.8, 0.16] output from step 1042 and the aggregated neighborhood node features of step 1043, such as [0.72, 0.12], are taken and concatenated into a new vector [0.8, 0.16, 0.72, 0.12]. Then, a nonlinear activation operation is applied, such as the ReLU function formula max(0, x), and each element is processed independently, such as inputting 4 elements and outputting 4 elements. For example, the first element 0.8 is taken, and after ReLU, it is still 0.8, and all elements are greater than 0, without change. The output is used as the trajectory decoupling feature, representing the separation feature path information of the target in the occlusion area. Finally, a feature vector such as [0.8, 0.16, 0.72, 0.12] is generated.

[0102] In practical applications, in a certain radar tracking scenario, faced with the signal occlusion problem caused by the intersection of trajectories of multiple flying targets, the scattering point data generated based on step 103, including the position, amplitude and occlusion relationship edges of 100 scattering points, is used to decouple trajectory features using dynamic graph convolution. In specific implementation, the amplitude value is first converted into a node feature: for example, the amplitude value of scattering point A is 3.2 units, which is converted into the initial feature vector [1.5, -0.3] through the fully connected layer. The feature is stored in memory as a 32-bit floating point number and occupies 8 bytes of space. Secondly, based on the dynamic edge weights calculated in step 103, such as the 4.01 unit between targets A and B, 2 / second is used as the neighborhood aggregation coefficient, and the initial features are linearly transformed: let the learnable weight matrix be a 2×2 matrix [[0.8,-0.2],[0.1,0.6]], and calculate the intermediate feature of point A = [1.5×0.8+(-0.3)×(-0.2),1.5×0.1+(-0.3)×0.6]=[1.26,-0.03]. Next, we filter the neighboring nodes: When point A is connected to points B and C via an occlusion edge (the intermediate feature of point B is [0.9, 0.4], and that of point C is [-0.5, 1.2]), we weight them using the edge weights: if the AB edge weight is 4.01, the weighted feature of point B = [0.9 × 4.01, 0.4 × 4.01] = [3.61, 1.60]; if the AC edge weight is 2.3, the weighted feature of point C = [-0.5 × 2.3, 1.2 × 2.3] = [-1.15, 2.76]. The aggregated neighborhood feature = [3.61 + (-1.15), 1.60 + 2.76] = [2.46, 4.36]. Finally, the aggregated features are concatenated with the intermediate features of point A itself [-0.03] to obtain [2.46, 4.36, -0.03]. After ReLU activation, the threshold is 0, and negative values ​​are reset to zero, the final trajectory decoupling feature [2.46, 4.36, 0] is output. This three-dimensional feature vector describes the motion characteristics of the target in the trajectory intersection area.

[0103] In the overall scheme of the above step 104, by innovatively integrating the scattering point amplitude information and the dynamic occlusion relationship features, the scattering point amplitude values ​​are first converted into initial node feature weights to achieve a preliminary characterization of the target's physical characteristics, and then the dynamic edge weights are accurately injected into the neighborhood aggregation coefficient of the graph attention mechanism, and the initial features are linearly transformed through the learnable weight matrix to generate intermediate node features containing spatial correlations; then, the neighborhood node set is dynamically screened based on the occlusion relationship edges, and the intermediate features of the neighborhood nodes are weighted and summed with the dynamic edge weights as coefficients, creatively realizing the fusion and transmission of occlusion area features; finally, by splicing the current node and neighborhood aggregation features and applying nonlinear activation, the spatial-temporal evolution characteristics of the target in the occlusion and trajectory intersection area are deeply explored, and a breakthrough trajectory decoupling feature with strong discriminability is generated. This feature not only retains the morphological structure information of the target in the occlusion state, but also effectively decouples the trajectory confusion effect, providing a key discrimination basis for accurate multi-target recognition in complex scenarios.

[0104] 105. Generate association weights based on the trajectory decoupling features, input into a multi-hypothesis tracking algorithm to generate multi-branch trajectory hypotheses, perform conflict resolution based on the instantaneous velocity vector angles of the multi-branch trajectory hypotheses, and when the spatial distance between two branches is less than a dynamic threshold and the velocity vector angle exceeds a set angle, retain the branch trajectory hypothesis with a higher association weight and output a multi-target trajectory sequence.

[0105] Optionally, step 105 may specifically include the following steps:

[0106] 1051. Based on the trajectory decoupling feature, calculate the similarity of feature vectors of different trajectory points, generate association weights, and input the association weights into a multi-hypothesis tracking algorithm to generate multi-branch trajectory hypotheses;

[0107] 1052. For the multi-branch trajectory hypothesis, when the spatial distance between any two branches is less than a dynamic threshold, further calculate the instantaneous velocity vector angle of the two branches. When the instantaneous velocity vector angle exceeds a set angle, compare the association weights of the two branches, retain the branch trajectory hypothesis with the higher association weight, integrate the retained branch trajectory hypotheses, and output the multi-target trajectory sequence in chronological order.

[0108] Among them, step 1052 may specifically include the following processes: screening the multi-branch trajectory hypotheses whose spatial distance at the current moment is less than the dynamic threshold to form a set of branch pairs to be processed, and extracting the instantaneous velocity vector of each branch trajectory pair in the set of branch pairs to be processed at the current moment; calculating the angle value of the two instantaneous velocity vectors in the same branch trajectory pair, and the angle value is obtained by vector dot product operation and arc cosine function conversion; when the angle value exceeds the set angle, extracting the association weight values ​​of the two branch trajectories from the set of branch pairs to be processed; comparing the association weight values ​​of the two branch trajectories, retaining the branch trajectory with a larger association weight value, and removing the branch trajectory with a smaller association weight value; integrating the retained branch trajectory hypotheses, and outputting a multi-target trajectory sequence in chronological order.

[0109] In the above steps, the trajectory decoupling feature is the feature vector data output from step 104 representing the independent movement of the target in space and time, including position and velocity information; the association weight is a value calculated by comparing the similarity of the feature vectors of different trajectory points, which represents the strength value of the correlation between trajectory points; the multi-hypothesis tracking algorithm is a standard tracking method that generates multiple possible target motion path schemes; the multi-branch trajectory hypothesis is a plurality of trajectory path branches generated by the algorithm, each branch represents a complete trajectory sequence of a potential target; the instantaneous velocity vector angle is the difference angle between the velocity directions of the two branches at the same time point, which is calculated by the vector dot product and the inverse cosine function; conflict resolution is the processing process when branches overlap, which is used to eliminate unreasonable trajectory paths; the dynamic threshold is the distance critical value that changes with time or scene, which is used to judge whether the distance between branches is small enough; the set angle is pre-set A defined angle standard value, such as 90 degrees, is used as the boundary for judging speed direction differences; a branch trajectory hypothesis with a higher association weight indicates a more reliable trajectory branch retained in conflict resolution; a multi-target trajectory sequence is the trajectory data of multiple targets in the time series that is finally output; a trajectory point is a single position point in the trajectory path; feature vector similarity is a quantitative value of the degree of difference between the feature vectors of trajectory points calculated by mathematical methods such as cosine similarity; a branch trajectory pair set is a pairing set consisting of multiple trajectory branches; a pending branch pair set is a set consisting of conflicting branch pairs that meet the distance condition at the current time point; an instantaneous speed vector is a direction quantity consisting of the speed direction and magnitude of a branch at a certain point in time; a vector dot product operation is a mathematical formula for multiplying and summing two vectors; an inverse cosine function is a mathematical function for calculating angle values; and an included angle value is a specific angle value calculated and used to judge speed direction differences.

[0110] In the embodiment of the present application, first, in step 1051, the similarity of the feature vectors of different trajectory points is calculated based on the trajectory decoupling feature to generate an associated weight, and the weight is input into the multi-hypothesis tracking algorithm to generate a multi-branch trajectory hypothesis; the specific process is: extract the feature vector of each trajectory point from the trajectory decoupling feature, such as the point feature of target 1 at time t1 is [0.8, 0.16, 0.72, 0.12], and calculate the similarity of any two points using the cosine similarity formula Among them F i and F jis the feature vector of track point i and j (unitless), the dot product is the sum of the multiplication of vector elements, and the length is the Euclidean norm of the vector, for example, point A feature [0.8, 0.16, 0.72, 0.12] and point B feature [0.9, 0.18, 0.75, 0.13], the dot product is 0.8*0.9+0.16*0.18+0.72*0.75+0.12*0.13=0.72+0.0288+0.54+0.0156≈1.3044, and the length is Similarly, ||F j ||≈1.115 Similarity = 1.3044 / (1.095*1.115)≈1.3044 / 1.221≈1.068, the normalized weight of 0.8 represents 0.8; then the similarity value is directly converted into the association weight, for example, the similarity of point A-B 1.068 is mapped to the weight value 0.8, and all point pair weights are combined into a matrix; then a multi-hypothesis tracking algorithm such as JPDA algorithm is input, which generates multiple track branch hypotheses based on the weight, for example, hypothesis track branch 1 represents the possible path of target A, and branch 2 represents the possible path of target B, and outputs a multi-branch track hypothesis sequence, for example, branch 1 position sequence [100 meters, 102 meters,...], and branch 2 sequence [110 meters, 112 meters,...]). For example, in a radar tracking scene, there are 10 track points, a weight matrix is generated by this calculation, and 3 branch track hypotheses are output after inputting the algorithm.

[0111] Secondly, conflict resolution is performed based on the multi-branch track hypothesis through step 1052, when the spatial distance of two branches is less than the dynamic threshold and the included angle of the velocity vectors exceeds the set angle, the association weight is compared and the branch track hypothesis with higher weight is retained, and a multi-target track sequence is integrated and output; the specific process is as follows: first, multiple branch track hypotheses with a spatial distance less than the dynamic threshold at the current time point are screened, for example, the dynamic threshold is calculated based on the radar resolution and speed, and the initial value is 0.5 meters, which is automatically adjusted over time, forming a set of branch pairs to be processed, for example, the distance between branch 1 and branch 2 at time t is |100 meters-102 meters|=2 meters, but the dynamic threshold is 2.5 meters, which does not meet the requirement; the distance between branch 3 and branch 4 is |150 meters-150.3 meters|=0.3 meters, which is less than the threshold 0.5 meters, and is added to the set ()); then, the instantaneous velocity vectors of the branch pairs in the set at the current time point are extracted (for example, the velocity vector of branch 3 at t is [20 meters / second, 5 meters / second], which represents a component of 20 meters / second direction and 5 meters / second amplitude); the included angle of the two velocity vectors of the same branch pair is calculated, which is v 1x v 2x +v 1y v 2y , length included angle formula Among them, the inverse cosine function cos -1 Output angle value (unit degree), for example, branch 3 velocity vector [20, 5] and branch 4 velocity vector [10, 0], dot product 20*10+5*0=200, modulus Angle θ = cos -1 (200 / (20.62×10))≈cos -1 (200 / 206.2)≈cos -1 (0.97)≈14 degrees is less than the set angle of 90 degrees. If the set angle is exceeded, such as the speed of branch 5 [0,30] and the speed of branch 6 [30,0], the dot product is 0, the modulus is 30, and the angle θ = cos-1(0) = 90 degrees, which is equal to the preset value of 90 degrees, then the associated weights are extracted from the weight matrix output in step 1051, such as the weight of branch 5 is 0.7 and the weight of branch 6 is 0.4. The weight values ​​are compared and the branch trajectory hypothesis with the higher weight is retained. For example, if the weight of branch 5 is 0.7>0.4, branch 5 is retained and the branch with the lower weight is removed. Finally, all retained branch hypotheses are integrated, such as retaining branches 1, 2, and 5, and the trajectory point sequence is organized in chronological order, such as the target 1 output sequence [t1 position 100 meters, t2 position 105 meters, ...] and the target 2 sequence [t1 position 110 meters, t2 position 115 meters, ...] as a multi-target trajectory sequence. For example, in a target occlusion scene, five branches conflict, and after this resolution, three final trajectories are output.

[0112] In a practical application, in a certain flying target tracking scenario, to address the conflict caused by the intersection of multiple target trajectories, multi-target trajectory decision making is performed based on the trajectory decoupling features output in step 104. For example, the feature vectors of target A are [2.1, 3.0], target B are [1.8, 2.9], and target C are [-0.5, 4.2]. First, feature similarity is calculated to generate association weights: the dot product of the feature vectors of targets A and B is 2.1 × 1.8 + 3.0 × 2.9 = 3.78 + 8.7 = 12.48, which is normalized to an association weight of 0.85; the dot product of targets A and C is 2.1 × (-0.5) + 3.0 × 4.2 = -1.05 + 12.6 = 11.55, with a weight of 0.78. These weights are input into the multi-hypothesis tracking algorithm to generate three initial trajectory branch hypotheses: branch 1 corresponds to target A, branch 2 corresponds to target B, and branch 3 corresponds to target C. Detect trajectory conflict at a time step of 50 seconds: the real-time spatial positions of branch 1 and branch 2 are (2050m, 3100m) and (2045m, 3095m), respectively. Less than the dynamic threshold of 10m. Calculate the instantaneous velocity vectors of the two branches: Branch 1 velocity Branch 2 speed Vector Angle The set angle is less than 60°, so two branches are reserved; the included angle between branch 1 and branch 3 is 78° (more than 60°) and the correlation weight 0.85>0.78, so branch 1 is reserved. After integration, three clear trajectory sequences are output, effectively solving the tracking conflict problem in the intersection area.

[0113] In the overall scheme of the above step 105, the correlation weight is generated by utilizing the trajectory decoupling feature, and is input into the multi-hypothesis tracking algorithm to generate multi-branch trajectory hypotheses; when facing the trajectory intersection or occlusion scene, the instantaneous speed vector angle is innovatively introduced as the core basis for judging the trajectory conflict - when the spatial distance of two branches is less than the dynamic adaptive threshold and the speed vector angle exceeds the set angle threshold, the system automatically triggers the branch screening mechanism based on the correlation weight, and accurately reserves the trajectory branch with higher feature matching degree; the whole conflict resolution process deeply integrates the three discriminators of spatial distance constraint, motion direction difference quantization and feature similarity, and finally outputs the multi-target trajectory sequence with strong spatiotemporal continuity, trajectory ambiguity resolution ability and dynamic scene adaptation ability, providing a robust solution for centimeter-level trajectory tracking of clustered targets in strong occlusion and high-density intersection environment.

[0114] The following is a complete embodiment for steps 101 to 105:

[0115] As shown in Figure 2 , in a certain air traffic monitoring scene, the radar system tracks the flight target group that is mutually occluded and trajectory intersected in the region, the system first continuously collects 10 seconds of echo signals at a pulse repetition frequency of 10 kHz, forming 100 radar data corresponding to pulses (10 kHz x 10 seconds = 100 pulses), each pulse is processed to generate 200 distance sampling points, and finally arranged in time sequence into a 100 row x 200 column radar echo two-dimensional matrix.

[0116] Signal separation is achieved by analyzing the characteristics of the target scattering points, for example, a certain scattering point position moves uniformly from distance unit 10 to 50 in 100 time steps, showing linear characteristics, or changes along an arc path from position 15 to 48, showing curved surface characteristics. For linearly distributed points, when target A is at position (35.0) units at step 50 and position (35.2) units at step 51, which satisfies continuous smooth change, it is classified as the same target; for curved surface distributed points, when the position change curvature difference value is less than 0.1 units at step 25 and step 50, and the position continuously increases throughout the process, it is determined as the same target, and this process finally outputs the three-dimensional positions of 100 scattering points, such as target A located at (30, 40, 50) units and its amplitude value 3.2 units.

[0117] Based on the separated data, the target relationship is constructed. At step 60, the calculated spatial distance between target A's position (32, 41, 53) and target B's position (31, 40, 52) is 1.73 units. Meanwhile, target A moves from (22, 32, 42) to (32, 42, 52) in 2 seconds at a calculated speed of approximately 8.66 units per second, while target B's speed is 7.8 units per second. The speed difference between the two is 0.86 units per second, resulting in a dynamic edge weight of 1.49 units squared per second. An occlusion relationship edge is established when the distance between the two targets remains less than 0.5 units between steps 55 and 65. The amplitude value of target A, 3.2 units, is converted to a feature vector [1.5, -0.3]. This is then transformed through a weight matrix to obtain an intermediate feature [1.26, -0.03]. Neighboring node features, such as the weighted feature of node B [1.34, 0.60], are then aggregated and concatenated to produce a trajectory decoupling feature [0.19, 3.36, 0] using an activation function.

[0118] Finally, a multi-target trajectory is generated. When the spatial distance between the unit at branch 1 (2105, 3150) and the unit at branch 2 (2100, 3145) is 7.07 meters, which is less than the 10-meter threshold: the velocity vector of branch 1 (15, 20) and the velocity vector of branch 2 (-5, 30), with an angle of 46 degrees less than the 60-degree threshold, are retained. When the angle between branch 1 and branch 3 exceeds the threshold of 78 degrees and the association weight of 0.85 is greater than 0.78, the high-weight branch is retained first, and the decoupled trajectory sequence is finally output.

[0119] Figure 3 The present invention provides a schematic diagram of a radar signal multi-target tracking system based on a dynamic graph convolutional network. Figure 3 As shown, the system includes:

[0120] An acquisition module 31 is configured to acquire a one-dimensional radar range profile sequence of a target cluster, perform pulse compression processing on the one-dimensional range profile sequence, and arrange the sequence in a slow time dimension to form a two-dimensional radar echo sequence;

[0121] A separation module 32 is configured to separate the positions and amplitudes of aliased scattering points in the two-dimensional sequence of radar echoes according to the linear or curved distribution characteristics of the scattering points of the cone target, thereby obtaining scattering point position and amplitude data;

[0122] A construction module 33 is configured to calculate the spatial distance between targets based on the position information in the scattering point position and the amplitude data, and construct a dynamic edge weight based on the radial velocity difference, and generate an occlusion relationship edge when the spatial distance between targets is less than a radar resolution threshold;

[0123] A generation module 34 is configured to initialize node feature weights based on the scattering point positions and amplitude information in the amplitude data, use the dynamic edge weights as neighborhood aggregation coefficients of a graph attention mechanism, aggregate neighborhood node features of the occlusion relationship edges through a dynamic graph convolutional network, learn decoupling features of the target trajectory in the occlusion and intersection regions, and generate trajectory decoupling features;

[0124] Output module 35 is used to generate association weights based on the trajectory decoupling features, input them into a multi-hypothesis tracking algorithm to generate multi-branch trajectory hypotheses, and perform conflict resolution based on the instantaneous velocity vector angles of the multi-branch trajectory hypotheses. When the spatial distance between the two branches is less than a dynamic threshold and the velocity vector angle exceeds a set angle, the branch trajectory hypothesis with a higher association weight is retained, and a multi-target trajectory sequence is output.

[0125] Figure 3 The radar signal multi-target tracking system based on dynamic graph convolutional network can be performed Figure 1 The implementation principles and technical effects of the radar signal multi-target tracking method based on a dynamic graph convolutional network described in the illustrated embodiment will not be elaborated on here. The specific manner in which each module and unit performs operations in the radar signal multi-target tracking system based on a dynamic graph convolutional network in the above embodiment has been described in detail in the embodiments of the method and will not be elaborated on here.

[0126] In one possible design, Figure 3 The radar signal multi-target tracking system based on dynamic graph convolutional network of the embodiment shown can be implemented as a computing device, such as Figure 4 As shown, the computing device may include a storage component 41 and a processing component 42;

[0127] The storage component 41 stores one or more computer instructions, wherein the one or more computer instructions are called and executed by the processing component 42 .

[0128] The processing component 42 is used for the above Figure 1 The embodiment provides a radar signal multi-target tracking method based on a dynamic graph convolutional network.

[0129] The processing component 42 may include one or more processors to execute computer instructions to perform all or part of the steps in the above method. Of course, the processing component may also be implemented as one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the above method.

[0130] The storage component 41 is configured to store various types of data to support operations at the terminal. The storage component can be implemented by any type of volatile or non-volatile memory device, or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk, or optical disk.

[0131] Of course, a computing device may also include other components, such as input / output interfaces, display components, communication components, etc.

[0132] The input / output interface provides an interface between the processing component and the peripheral interface module, which can be an output device, an input device, etc.

[0133] The communication component is configured to facilitate, among other things, wired or wireless communications between the computing device and other devices.

[0134] Among them, the computing device can be a physical device or an elastic computing host provided by a cloud computing platform, etc. In this case, the computing device can refer to a cloud server, and the above-mentioned processing components, storage components, etc. can be basic server resources rented or purchased from the cloud computing platform.

[0135] The present application also provides a computer storage medium storing a computer program, wherein the computer program can achieve the above-mentioned Figure 1 The illustrated embodiment provides a radar signal multi-target tracking method based on a dynamic graph convolutional network.

[0136] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0137] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e., they may be located in one location or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of the present embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0138] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, or of course, by means of hardware. Based on this understanding, the essence of the above technical solution or the part that contributes to the prior art can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions for enabling a computer device (which can be a personal computer, a server, etc.) to execute the methods described in each embodiment or certain parts of the embodiments.

[0139] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A radar signal multi-target tracking method based on dynamic graph convolutional network, characterized in that: include: Collecting a one-dimensional radar range image sequence of the target cluster, performing pulse compression processing on the one-dimensional range image sequence, and arranging it in a slow time dimension to form a two-dimensional radar echo sequence; Separating the positions and amplitudes of aliased scattering points in the two-dimensional radar echo sequence according to the linear or curved distribution characteristics of the scattering points of the cone target to obtain scattering point position and amplitude data; Based on the position information in the scattering point position and amplitude data, the spatial distance between targets is calculated, and the dynamic edge weight is constructed in combination with the radial velocity difference. When the spatial distance between targets is less than the radar resolution threshold, an occlusion relationship edge is generated; Initialize node feature weights based on the scattering point positions and amplitude information in the amplitude data, use the dynamic edge weights as neighborhood aggregation coefficients of the graph attention mechanism, aggregate neighborhood node features of the occlusion relationship edges through a dynamic graph convolutional network, learn the decoupling features of the target trajectory in the occlusion and intersection areas, and generate trajectory decoupling features; Based on the trajectory decoupling features, an association weight is generated and input into a multi-hypothesis tracking algorithm to generate multi-branch trajectory hypotheses. Conflict resolution is performed based on the instantaneous velocity vector angles of the multi-branch trajectory hypotheses. When the spatial distance between the two branches is less than a dynamic threshold and the velocity vector angle exceeds a set angle, the branch trajectory hypothesis with a higher association weight is retained, and a multi-target trajectory sequence is output.

2. The method according to claim 1, characterized in that Initialize node feature weights based on the scattering point positions and amplitude information in the amplitude data, use the dynamic edge weights as the neighborhood aggregation coefficient of the graph attention mechanism, aggregate neighborhood node features of the occlusion relationship edges through a dynamic graph convolutional network, learn the decoupling features of the target trajectory in the occlusion and intersection areas, and generate trajectory decoupling features, including: Based on the scattering point position and amplitude information in the amplitude data, the amplitude value is converted into an initial node feature vector, and the node feature weight is initialized based on the initial node feature vector; Assigning the dynamic edge weight to the neighborhood aggregation coefficient of the graph attention mechanism, performing a linear transformation operation on the initial node feature vector through a learnable weight matrix to generate intermediate node features; Through a dynamic graph convolutional network, taking the occlusion relationship edge as the connection path, the features of the intermediate nodes in the neighborhood of the occlusion relationship edge are weighted summed to aggregate the neighborhood node features; The neighborhood node features after weighted summation are spliced ​​with the intermediate node features of the current node, and a nonlinear activation operation is performed on the spliced ​​features to learn the decoupling features of the target trajectory in the occlusion and intersection areas, and output the updated feature vector as the trajectory decoupling feature.

3. The method according to claim 1, characterized in that Based on the position information of the scattering point position and amplitude data, the spatial distance between targets is calculated, and the dynamic edge weight is constructed in combination with the radial velocity difference. When the spatial distance between targets is less than the radar resolution threshold, an occlusion relationship edge is generated, including: Calculating the inter-target spatial distance between any two scattering points based on the position information in the scattering point positions and amplitude data, and calculating the radial velocity difference of the scattering points based on the change of the position information in the slow time dimension; Performing a product operation on the spatial distance between the targets and the radial velocity difference, and constructing a dynamic edge weight based on the product operation result; When the spatial distance between the targets is less than the radar resolution threshold, an occlusion relationship edge is generated between the corresponding scattering points, and the dynamic edge weight is used as the association strength value of the occlusion relationship edge.

4. The method according to claim 1, wherein Generate association weights based on the trajectory decoupling features, input into a multi-hypothesis tracking algorithm to generate multi-branch trajectory hypotheses, perform conflict resolution based on the instantaneous velocity vector angles of the multi-branch trajectory hypotheses, retain the branch trajectory hypothesis with the higher association weight when the spatial distance between the two branches is less than a dynamic threshold and the velocity vector angle exceeds a set angle, and output a multi-target trajectory sequence, including: Based on the trajectory decoupling feature, calculating the similarity of feature vectors of different trajectory points, generating association weights, and inputting the association weights into a multi-hypothesis tracking algorithm to generate multi-branch trajectory hypotheses; For the multi-branch trajectory hypothesis, when the spatial distance between any two branches is less than a dynamic threshold, the instantaneous velocity vector angle of the two branches is further calculated. When the instantaneous velocity vector angle exceeds the set angle, the association weights of the two branches are compared, and the branch trajectory hypothesis with the higher association weight is retained. The retained branch trajectory hypotheses are integrated and the multi-target trajectory sequence is output in chronological order.

5. The method according to claim 4, characterized in that For the multi-branch trajectory hypothesis, when the spatial distance between any two branches is less than a dynamic threshold, the instantaneous velocity vector angle of the two branches is further calculated. When the instantaneous velocity vector angle exceeds a set angle, the association weights of the two branches are compared, and the branch trajectory hypothesis with the higher association weight is retained. The retained branch trajectory hypotheses are integrated and the multi-target trajectory sequence is output in chronological order, including: Filtering the multi-branch trajectory hypotheses whose spatial distance at the current moment is less than a dynamic threshold to form a set of branch pairs to be processed, and extracting the instantaneous velocity vector of each branch trajectory pair in the set of branch pairs to be processed at the current moment; Calculate the angle between two instantaneous velocity vectors in the same branch trajectory pair, where the angle is obtained by vector dot product operation and arc cosine function conversion; When the angle value exceeds the set angle, extract the associated weight values ​​of two branch trajectories from the set of branch pairs to be processed; compare the associated weight values ​​of the two branch trajectories, retain the branch trajectory with a larger associated weight value, and remove the branch trajectory with a smaller associated weight value; The retained branch trajectory hypotheses are integrated and a multi-target trajectory sequence is output in chronological order.

6. The method according to claim 1, characterized in that Separating the positions and amplitudes of aliased scattering points in the two-dimensional radar echo sequence according to the linear or curved distribution characteristics of the scattering points of the cone target, and obtaining scattering point position and amplitude data, including: Extracting a change sequence of the position information of each scattering point in the two-dimensional radar echo sequence in a slow time dimension, analyzing a continuous offset pattern of the change sequence, and identifying whether the position distribution characteristic of the change sequence is a linear distribution characteristic or a curved surface distribution characteristic; With respect to the linear distribution characteristics of the scattering points of the cone target, the scattering points that meet the continuity condition are classified into the same target according to the continuity of the position offset direction of adjacent slow time frames and the smoothness of the amplitude change; Calculating a curvature change value of a position change sequence of the scattering points of the cone target based on the surface distribution characteristics of the scattering points. When the curvature change value is less than a preset threshold and the position offset amplitude continues to increase, determining that the scattering points belong to the same target. Based on the scattering points classified into the same target, the position information of the overlapping spatial positions is detected, the aliased scattering point positions and amplitudes are separated, and the separated scattering point position and amplitude data are obtained.

7. The method according to claim 2, characterized in that Through a dynamic graph convolutional network, with the occlusion relationship edge as the connection path, the features of the intermediate nodes in the neighborhood of the occlusion relationship edge are weighted summed to aggregate the neighborhood node features, including: Based on the occlusion relationship edge, all neighboring nodes that the current node is directly connected to through the occlusion relationship edge are filtered out to form a neighboring node set; Extracting intermediate node features of each node in the neighborhood node set, wherein the intermediate node features are generated by linearly transforming the initial node features; The dynamic edge weight is used as a neighborhood aggregation coefficient, and the intermediate node feature of each neighborhood node is multiplied by the corresponding neighborhood aggregation coefficient to generate a weighted neighborhood node feature; A sum operation is performed on all the weighted neighborhood node features of the same current node, and an aggregated neighborhood node feature is output.

8. A radar signal multi-target tracking system based on dynamic graph convolutional network, characterized in that: include: An acquisition module is used to acquire a one-dimensional radar range image sequence of a target cluster, perform pulse compression processing on the one-dimensional range image sequence, and arrange it in a slow time dimension to form a two-dimensional radar echo sequence; A separation module is used to separate the positions and amplitudes of aliased scattering points in the two-dimensional sequence of radar echoes according to the linear or curved distribution characteristics of the scattering points of the cone target, so as to obtain scattering point position and amplitude data; A construction module is configured to calculate the spatial distance between targets based on the position information in the scattering point position and amplitude data, and to construct a dynamic edge weight based on the radial velocity difference, and to generate an occlusion relationship edge when the spatial distance between targets is less than a radar resolution threshold; A generation module is configured to initialize node feature weights based on the scattering point positions and amplitude information in the amplitude data, use the dynamic edge weights as neighborhood aggregation coefficients of a graph attention mechanism, aggregate neighborhood node features of the occlusion relationship edges through a dynamic graph convolutional network, learn the decoupling features of the target trajectory in the occlusion and intersection areas, and generate trajectory decoupling features; The output module is used to generate association weights based on the trajectory decoupling features, input the multi-hypothesis tracking algorithm to generate multi-branch trajectory hypotheses, and perform conflict resolution based on the instantaneous velocity vector angles of the multi-branch trajectory hypotheses. When the spatial distance between the two branches is less than a dynamic threshold and the velocity vector angle exceeds a set angle, the branch trajectory hypothesis with a higher association weight is retained, and a multi-target trajectory sequence is output.

9. A computing device, characterized in that It includes a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are used to be called and executed by the processing component to implement a radar signal multi-target tracking method based on a dynamic graph convolutional network as described in any one of claims 1 to 7.

10. A computer storage medium, characterized in that A computer program is stored, and when the computer program is executed by a computer, the method for multi-target tracking of radar signals based on a dynamic graph convolutional network as claimed in any one of claims 1 to 7 is implemented.

Images