A radar multi-target point cloud registration method

Abstract

The application discloses a radar multi-target point cloud registration method, and relates to the technical field of millimeter wave radar sensing and intelligent signal processing. The method comprises the following steps: acquiring radar echo signals according to frames, and constructing a space-time point cloud signal; adopting CFAR detection to extract space-time units with signal strength higher than a threshold as potential target point cloud areas; dividing the potential areas into a plurality of target point cloud clusters, setting initial clustering parameters, introducing a density factor and a distance weight to determine cluster centers in an iteration process, constructing a comprehensive evaluation index and adaptively adjusting clustering parameters, eliminating noise clusters, and obtaining target point clouds in the multi-target point cloud clusters; and adopting an iterative closest point algorithm to register adjacent frame point clouds. The application can improve the multi-target point cloud separation precision and the inter-frame registration stability.

Description

Technical Field

[0001] This application relates to the field of millimeter-wave radar sensing and intelligent signal processing technology, and in particular to a radar multi-target point cloud registration method. Background Technology

[0002] Millimeter-wave radar boasts advantages such as all-weather operation, resistance to light interference, minimal impact on privacy, and sensitivity to moving targets. In recent years, it has been widely applied in scenarios such as indoor behavior perception, multi-target detection, and trajectory tracking. In indoor environments with multiple moving targets, radar echoes typically exhibit characteristics such as a large number of targets, significant echo overlap, strong clutter interference, and frequent obstruction and intersection. This makes stable separation of multi-target point clouds, inter-frame registration, and continuous tracking critical issues in engineering implementation.

[0003] In existing technologies, the registration of radar point clouds often employs the DBSCAN clustering method based on Euclidean distance or the clustering method based on simple centroid clustering to divide potential target point clouds into clusters, thereby achieving registration of target point clouds within clusters. However, the aforementioned existing technologies rely on preset fixed clustering parameters for point cloud cluster division. These fixed clustering parameters cannot be adapted to special situations such as uneven point cloud density, large changes in target shape, or strong scene noise. This can lead to issues such as mismerging of near-range target point clouds and over-segmentation of sparse targets, resulting in insufficient point cloud registration accuracy. Summary of the Invention

[0004] Therefore, it is necessary to provide a radar multi-target point cloud registration method to address the aforementioned technical problems.

[0005] The present invention adopts the following technical solution: This invention provides a radar multi-target point cloud registration method, comprising: The radar echo signals from multiple moving targets are acquired frame by frame; and a spatiotemporal point cloud signal for each frame is constructed based on the radar echo signals. The CFAR detection method is used to identify the region of spatiotemporal point cloud signal with a signal strength higher than a preset signal strength threshold in each frame as the potential target point cloud region for each frame. The potential target point cloud region of each frame is divided into several target point cloud clusters, and initial clustering parameters are set to perform clustering processing on the target point clouds in each target point cloud cluster. During the iterative process of clustering processing, the radar echo signal intensity corresponding to the target point cloud is introduced as a range weight to determine the centroid position of each target point cloud cluster. Based on the centroid position of the target point cloud cluster, a comprehensive evaluation index is constructed to assess the rationality of point cloud assignment. With the goal of minimizing the comprehensive evaluation index, the initial clustering parameters are adaptively adjusted to obtain the point cloud redistribution result in each target point cloud cluster. Based on the point cloud redistribution result in each target point cloud cluster, multi-target point cloud clusters in the potential target point cloud region of each frame are determined. The iterative nearest point algorithm is used to register the target point clouds in the multi-target point cloud clusters in the potential target point cloud region of adjacent frames, so as to realize the multi-target point cloud registration of the radar.

[0006] Preferably, based on the radar echo signal, a spatiotemporal point cloud signal is constructed, specifically including: Two-dimensional fast Fourier transforms are performed on the radar echo signal in the range and velocity dimensions respectively to obtain the range-Doppler matrix; range and velocity parameters of potential target echoes are extracted from the range-Doppler matrix by peak search; For the peak values ​​of the range and velocity parameters of potential target echoes, array snapshot data corresponding to the potential target echoes within the range gate is extracted. Based on the array snapshot data, a covariance matrix of the echo received signal is constructed. A diagonal loading technique is introduced into the covariance matrix of the echo received signal to robustly optimize the covariance matrix. According to the robustly optimized covariance matrix, a steering vector matching the array snapshot data structure is constructed. Based on the steering vector and the robustly optimized covariance matrix, the spatial power spectral density function is determined using the Capon minimum variance criterion. A global peak search is performed on the spatial power spectral density function to obtain the azimuth information of the potential target echo. The range parameters, velocity parameters, azimuth information, and frame sequence information of the potential target echo are fused to obtain a spatiotemporal point cloud signal.

[0007] Preferably, the initial clustering parameters include: initial cluster centers, lateral distance threshold, longitudinal distance threshold, velocity threshold, iteration number threshold, density threshold, and convergence condition.

[0008] Preferably, target point cloud distance weights are introduced to determine the centroid position of each target point cloud cluster, specifically including: Transform the target point clouds in each target point cloud cluster from the polar coordinate system to the rectangular coordinate system to obtain the lateral distance, longitudinal distance and velocity of each target point cloud; The target point clouds whose lateral distance is within the lateral distance threshold are merged to obtain a first target point cloud set; the lateral distance of each target point cloud in the first target point cloud set is weighted by the point cloud amplitude of each target point cloud in the first target point cloud set, and the lateral distance of each target point cloud in the first target point cloud set is weighted to obtain the target point cloud lateral distance of each target point cloud cluster. Merge the target point clouds in the first target point cloud set that are within the longitudinal distance threshold to obtain the second target point cloud set; use the point cloud amplitude of each target point cloud in the second target point cloud set as the longitudinal distance weight, and perform a weighted average of the longitudinal distances of each target point cloud in the second target point cloud cluster set to obtain the longitudinal distance of the target point cloud in each target point cloud cluster. The target point clouds within the velocity threshold in the second target point cloud set are merged to obtain the third target point cloud set; the velocity of each target point cloud in the third target point cloud set is weighted by the point cloud amplitude of each target point cloud in the third target point cloud set to obtain the target point cloud velocity of each target point cloud cluster. The horizontal distance and vertical distance of the target point cloud are determined as the centroid positions of the target point cloud cluster, and each centroid position corresponds to the instantaneous velocity of the target point cloud cluster.

[0009] Preferably, a comprehensive evaluation index for assessing the rationality of point cloud attribution is constructed based on the centroid location of the target point cloud region, specifically including: For each target point cloud in a target point cloud cluster, determine the centroid position of the target point cloud and its own target point cloud cluster, as well as the cross-entropy between the target point cloud and the centroid positions of its neighboring target point cloud clusters; The global density of each target point cloud cluster is determined based on the local density of each target point cloud in each target point cloud cluster. Based on the cross-entropy and the global density of the target point cloud cluster, a comprehensive evaluation index is constructed to assess the rationality of point cloud attribution. Preferably, the local density of the target point cloud in the target point cloud cluster is determined by the ratio of the number of point clouds in the target point cloud within a preset range to the total number of target point clouds in the target point cloud cluster; the global density is determined by summing the local densities of each target point cloud in each target point cloud cluster and then dividing by the total number of target point cloud clusters.

[0010] Preferably, the iterative nearest-neighbor algorithm is used to register the target point clouds in the multi-target point cloud clusters in the potential target point cloud regions of adjacent frames, specifically including: The geometric center of the multi-target point cloud cluster in the potential target point cloud region of the current frame is selected as the registration point set, and the geometric center of the multi-target point cloud cluster after processing in the previous frame is selected as the reference point set. The nearest distance between each point in the set of points to be registered and each point in the set of reference points is calculated based on Euclidean distance, and an initial set of point pairs is established. The azimuth information in the spatiotemporal point cloud signal is used as the angle constraint weight to correct the nearest distance in the initial set of point pairs, and a corrected set of point pairs is obtained. Based on the centroid positions of the multi-target point cloud clusters in the potential target point cloud regions of the current frame and the previous frame, as well as the azimuth information in the spatiotemporal point cloud signal, an attitude error objective function is constructed. By performing singular value decomposition on the set of correction point pairs, the optimal rotation and translation matrices that minimize the attitude error objective function are solved. The optimal rotation and translation matrices are then used to perform coordinate transformation on the target point clouds in the multi-target point cloud clusters of the current frame, thereby realizing the target point clouds in the multi-target point cloud clusters in the potential target point cloud regions of adjacent frames.

[0011] Preferably, after registering the target point clouds in the multi-target point cloud clusters of adjacent frames, the method further includes: A multi-unscented Kalman filter tracker is constructed based on the SORT framework. The target point cloud state vector is formed by the centroid position coordinates, the projection components of the Doppler velocity in three-dimensional space, and the azimuth and pitch information. Using the target point cloud state vector, state covariance matrix, and Sigma point extension parameters as input parameters for the constant velocity model, the system outputs the target state value and its covariance matrix of the target point cloud in the current frame. Based on the actual detected point cloud state value as the observation value, the system performs unscented Kalman filtering to update the target point cloud state estimate and outputs the smooth trajectory points of the target point cloud in the current frame.

[0012] Preferably, after outputting the smooth trajectory of the target point cloud in the multi-target point cloud cluster of the current frame, the method further includes: The target state estimate of the next frame target point cloud output by the multi-unscented Kalman filter tracker is the tracking threshold center. The Mahalanobis distance between the tracking threshold center and the centroid position of the current frame target point cloud is calculated. A joint cost matrix is ​​constructed based on the Mahalanobis distance between the tracking threshold center and the centroid position of the target point cloud in the current frame, the centroid position deviation term between the tracking threshold and the target point cloud in the current frame, the velocity consistency term, and the angle consistency term, and the associated threshold parameters are set. The joint cost matrix is ​​input into the Hungarian algorithm to solve for the global optimal point cloud matching pair between the smooth trajectory of the target point cloud in the current frame and the point cloud trajectory in the historical frames.

[0013] Preferably, after solving the global optimal point cloud matching pairs between the smoothed trajectories of the target point clouds in the current frame's multi-target point cloud cluster and the historical trajectories, the method further includes: For successfully matched point cloud pairs, when the number of consecutive successful matches exceeds the confirmation threshold, the corresponding point cloud trajectory is marked as a confirmed point cloud trajectory and output. For confirmed trajectories that have not been successfully matched, the memory mechanism is activated to mark them as pending confirmation and increment the loss counter to record the state information of the target point cloud before it is lost. If the target point cloud trajectory fails to detect a matching point cloud again within the preset loss count threshold, or if the predicted position of the target point cloud exceeds the radar field of view, it is determined that the target point cloud has left the scene, the corresponding trajectory is terminated, and the corresponding trajectory data is deleted.

[0014] The present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the aforementioned radar multi-target point cloud registration method.

[0015] The present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the aforementioned radar multi-target point cloud registration method.

[0016] The above-mentioned at least one technical solution adopted in this invention can achieve the following beneficial effects: The introduction of distance weighting means that the calculation of centroid position no longer relies solely on the average of spatial coordinates, but is weighted by spatial distance. This allows for accurate characterization of the true center of the cluster even when the target shape changes or the density is uneven. The comprehensive evaluation index constructed based on centroid position integrates the reasonableness of the attribution between the target point cloud and the target point cloud cluster center, as well as the global density distribution of adjacent clusters. It can dynamically distinguish between dense point clouds that should belong to the same target and neighboring target point clouds that should be separated, thereby avoiding over-merging or over-segmentation caused by fixed thresholds and improving the accuracy of point cloud registration in effective multi-target scenarios. Attached Figure Description

[0017] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0018] Figure 1 A flowchart illustrating a radar multi-target point cloud registration method provided by the present invention; Figure 2 This is a schematic diagram of the process of static clutter suppression, two-dimensional fast Fourier transform, and range-Doppler processing of the original echo data of the radar multi-target point cloud registration method provided by the present invention.

[0019] Figure 3This is a schematic diagram illustrating the process of candidate target extraction based on Capon spectral estimation and CFAR detection in a radar multi-target point cloud registration method provided by the present invention.

[0020] Figure 4 A flowchart illustrating the radar multi-target point cloud registration method provided by this invention, based on improved centroid clustering, ICP registration, multi-unscented Kalman filter tracking, and trajectory management; Figure 5 A diagram of a computer device for implementing a radar multi-target point cloud registration method provided by the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments in the specification without creative effort are within the scope of protection of this application.

[0022] In existing technologies, conventional range-Doppler processing can perform preliminary target detection. However, when multiple targets are close to each other, have similar velocities, or have insufficient angular resolution, relying solely on conventional beamforming or low-resolution angle estimation can easily lead to target aliasing, an increase in false spectral peaks, and insufficient angle positioning accuracy. Furthermore, directly performing fixed-threshold clustering or simple centroid calculation on the detection results often results in mis-merging of close-range targets, over-segmentation of sparse targets, and misattribution of noise points when point cloud density is uneven, target shape varies significantly, or scene noise is strong.

[0023] Furthermore, during target cross-frame association, if only a simple matching strategy based on geometric distance is used, trajectory switching and identity drift are prone to occur when targets cross, experience short-term occlusion, or are temporarily lost. Traditional linear state estimation methods have limited adaptability to nonlinear target state modeling that includes angle and Doppler information, making it difficult to balance trajectory smoothness, real-time performance, and robustness. Therefore, it is necessary to propose a radar multi-target point cloud separation and registration method that can simultaneously improve angle estimation accuracy, point cloud separation accuracy, inter-frame registration stability, and trajectory continuity.

[0024] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0025] Figure 1 This is a schematic diagram of a radar multi-target point cloud registration method according to the present invention, which specifically includes the following steps: S101: Acquire radar echo signals from multiple moving targets frame by frame; and construct a spatiotemporal point cloud signal for each frame based on the radar echo signals.

[0026] Specifically, raw echo data is acquired frame by frame and converted into a three-dimensional echo matrix. The three-dimensional matrix is ​​in the form of chirps×samples×channel, where chirps represents the number of pulses transmitted per frame, samples represents the number of sampling points per chirp, and channel represents the number of radar echo signal receiving antenna channels. Static clutter suppression processing is performed on the three-dimensional echo matrix to obtain the radar echo signal.

[0027] Optionally, based on the radar echo signal, a spatiotemporal point cloud signal is constructed, specifically including: performing a two-dimensional fast Fourier transform on the radar echo signal in both the range and velocity dimensions to obtain a range-Doppler matrix; extracting the range and velocity parameters of the potential target echo from the range-Doppler matrix through peak search; extracting array snapshot data of the corresponding potential target echo within a range gate based on the peak values ​​of the range and velocity parameters of the potential target echo; constructing a covariance matrix of the echo received signal based on the array snapshot data; and in the covariance matrix of the echo received signal... A diagonal loading technique is introduced to robustly optimize the covariance matrix of the received echo signal. Based on the robustly optimized covariance matrix, a steering vector matching the array snapshot data structure is constructed. Using the steering vector and the robustly optimized covariance matrix, the Capon minimum variance criterion is applied to determine the spatial power spectral density function. A global peak search is performed on the spatial power spectral density function to obtain the azimuth information of the potential target echo. The range parameters, velocity parameters, azimuth information, and frame sequence information of the potential target echo are fused to obtain the spatiotemporal point cloud signal.

[0028] S102: The CFAR detection method is used to identify the region of spatiotemporal point cloud signal with a signal strength higher than a preset signal strength threshold in each frame as the potential target point cloud region for each frame.

[0029] S103: Divide the potential target point cloud region of each frame into several target point cloud clusters, and set initial clustering parameters to perform clustering processing on the target point clouds in each target point cloud cluster; wherein, in the iterative process of clustering processing, the radar echo signal intensity corresponding to the target point cloud is introduced as a range weight to determine the centroid position of each target point cloud cluster, and a comprehensive evaluation index for evaluating the rationality of point cloud attribution is constructed based on the centroid position of the target point cloud cluster; with the goal of minimizing the comprehensive evaluation index, the initial clustering parameters are adaptively adjusted to obtain the point cloud redistribution result in each target point cloud cluster; based on the point cloud redistribution result in each target point cloud cluster, the multi-target point cloud clusters in the potential target point cloud region of each frame are determined.

[0030] Optionally, the initial clustering parameters include: initial cluster centers, lateral distance threshold, longitudinal distance threshold, velocity threshold, iteration number threshold, density threshold, and convergence condition.

[0031] Optionally, target point cloud distance weights are introduced to determine the centroid location of each target point cloud cluster, specifically including: The target point clouds in each target point cloud cluster are transformed from polar coordinates to Cartesian coordinates to obtain the lateral distance, longitudinal distance, and velocity of each target point cloud. Target point clouds whose lateral distance falls within the stated lateral distance threshold are merged to obtain a first target point cloud set. Using the point cloud amplitude of each target point cloud in the first target point cloud set as the lateral distance weight, a weighted average of the lateral distances of all target point clouds in the first target point cloud set is calculated to obtain the lateral distance of the target point clouds in each target point cloud cluster. Target point clouds in the first target point cloud set that fall within the longitudinal distance threshold are merged to obtain a second target point cloud set. The lateral distance of each target point cloud in the second target point cloud set is then calculated using the amplitude of the point cloud in the first target point cloud set as the lateral distance weight. The longitudinal distance weights are used to calculate the longitudinal distances of the target point clouds in the second target point cloud cluster set. The longitudinal distances of the target point clouds in each target point cloud cluster are then weighted and averaged to obtain the longitudinal distance of the target point cloud in each target point cloud cluster. The target point clouds in the second target point cloud set that are within the velocity threshold are merged to obtain the third target point cloud set. The velocity weights are used to calculate the velocity of the target point clouds in the third target point cloud set. The lateral distance and longitudinal distance of the target point clouds are determined as the centroid positions of the target point cloud clusters, and each centroid position corresponds to the velocity of the target point cloud cluster.

[0032] Optionally, a comprehensive evaluation index for assessing the rationality of point cloud attribution is constructed based on the centroid position of each target point cloud cluster. Specifically, this includes: for each target point cloud in a target point cloud cluster, determining the cross-entropy between the target point cloud and its own target point cloud cluster, and between the target point cloud and its neighboring target point cloud clusters; determining the global density of each target point cloud cluster based on the local density of each target point cloud in each target point cloud cluster; and constructing a comprehensive evaluation index for assessing the rationality of point cloud attribution based on the cross-entropy and the global density of the target point cloud cluster.

[0033] The local density of the target point cloud in the target point cloud cluster is determined by the ratio of the number of point clouds in the target point cloud within a preset range to the total number of target point clouds in the target point cloud cluster; the global density is determined by summing the local densities of each target point cloud in each target point cloud cluster and then dividing by the total number of target point cloud clusters.

[0034] S104: The iterative nearest point algorithm is used to register the target point clouds in the multi-target point cloud clusters in the potential target point cloud regions of adjacent frames, so as to realize the multi-target point cloud registration of the radar.

[0035] Optionally, an iterative nearest-point algorithm is used to register the target point clouds in the multi-target point cloud clusters in the potential target point cloud regions of adjacent frames. Specifically, this includes: selecting the geometric center of the multi-target point cloud clusters in the potential target point cloud region of the current frame as the set of points to be registered, and selecting the geometric center of the multi-target point cloud clusters processed in the previous frame as the set of reference points; calculating the nearest distance between each point in the set of points to be registered and each point in the set of reference points based on Euclidean distance, and establishing an initial set of point pairs; and using the azimuth information in the spatiotemporal point cloud signal as an angle constraint weight to correct the nearest distance in the initial set of point pairs. A set of corrected point pairs is obtained; based on the centroid positions of the multi-target point cloud clusters in the potential target point cloud regions of the current frame and the previous frame, and the azimuth information in the spatiotemporal point cloud signal as angular constraint weights, an attitude error objective function is constructed. By performing singular value decomposition on the set of corrected point pairs, the optimal rotation matrix and translation matrix that minimize the attitude error objective function are solved; the optimal rotation matrix and translation matrix are used to perform coordinate transformation on the target point clouds in the multi-target point cloud clusters of the current frame, thereby realizing the target point clouds in the multi-target point cloud clusters in the potential target point cloud regions of adjacent frames.

[0036] After registering the target point clouds in the multi-target point cloud clusters of adjacent frames, the process also includes: constructing a multi-unscented Kalman filter tracker based on the SORT framework, using the centroid position coordinates of the target point cloud, the projection components of the Doppler velocity in three-dimensional space, and the azimuth and pitch information to form the target point cloud state vector; combining the process noise covariance matrix, the observation noise covariance matrix, the initial error covariance matrix, and the Sigma point extension parameters, using a constant velocity model to predict the target state value and its covariance matrix of the target point cloud in the current frame, and using the actual detected point cloud state value as the observation value to perform unscented Kalman filtering update to obtain the target point cloud state estimate, and outputting the smooth trajectory points of the target point cloud in the current frame.

[0037] After outputting the smooth trajectory of the target point cloud in the current frame's multi-target point cloud cluster, the method further includes: using the target state estimate of the next frame's target point cloud output by a multi-unscented Kalman filter tracker as the tracking threshold center, calculating the Mahalanobis distance between the tracking threshold center and the centroid position of the current frame's target point cloud; constructing a joint cost matrix based on the Mahalanobis distance between the tracking threshold center and the centroid position of the current frame's target point cloud, the centroid position deviation term between the tracking threshold and the current frame's target point cloud, the velocity consistency term, and the angle consistency term, and setting the associated threshold parameter; inputting the joint cost matrix into the Hungarian algorithm to solve for the global optimal point cloud matching pair between the smooth trajectory of the current frame's target point cloud and the trajectory of the historical frame's point cloud.

[0038] After solving the global optimal point cloud matching pairs between the smooth trajectory of the target point cloud in the current frame's multi-target point cloud cluster and the historical trajectory, the process also includes: for successfully matched point cloud matching pairs, when the number of consecutive successful matches exceeds the confirmation threshold, the corresponding point cloud trajectory is marked as a confirmed point cloud trajectory and output; for confirmed trajectories that have not been successfully matched, a memory mechanism is activated to mark them as pending confirmation and increment the loss counter to record the state information of the target point cloud before it is lost; when the target point cloud trajectory has not detected a matching point cloud again within the preset loss count threshold, or when the predicted position of the target point cloud exceeds the radar field of view, it is determined that the target point cloud has left the scene, the corresponding trajectory is terminated, and the corresponding trajectory data is deleted.

[0039] The present invention provides specific embodiments to further illustrate the technical solutions provided above.

[0040] This embodiment presents a radar multi-target point cloud separation and registration method that integrates Capon spectrum estimation and improved centroid clustering. The overall process includes data acquisition and range-Doppler processing, angle estimation, candidate target detection, point cloud clustering and separation, point cloud registration, state estimation, trajectory association and lost target management.

[0041] In this embodiment, the main operating parameters of the millimeter-wave radar can be set as shown in Table 1.

[0042] Table 1. Examples of Radar Parameter Settings Raw Data Acquisition and Range-Doppler Processing: Raw echo data from an indoor multi-person movement scenario is acquired using millimeter-wave radar, and the raw echo data is reconstructed frame by frame. Static clutter suppression is applied to the reconstructed data to remove DC components and static background echoes. Subsequently, a two-dimensional fast Fourier transform is performed along the range and velocity dimensions to construct a range-Doppler matrix, and the range and velocity parameters of potential targets are obtained through peak search. The millimeter-wave radar can be an IWR1843BOOST radar platform. After format conversion, the raw echo data is formed into a three-dimensional matrix in the form of chirps×samples×channel, and static clutter is filtered out using phasor mean cancellation to enhance the dynamic target echo.

[0043] Specifically, the raw echo data is first converted into a three-dimensional matrix in the form of chirps × samples × channel, where chirps represents the number of pulses transmitted per frame, samples represents the number of sampling points per chirp, and channel represents the number of receiving antenna channels. Then, static clutter suppression processing is performed on the three-dimensional matrix to remove the effects of DC components and static background reflections.

[0044] In a preferred embodiment, a phasor mean cancellation algorithm can be used to process the phasor sequence within each distance gate. Let... The number of chirps per frame. For the echo phasor sequence within a single range gate, For phasor mean, For the dynamic target echo after clutter filtering, its mean and dynamic echo can be expressed as follows:

[0045] ; The above processing effectively suppresses fixed background and DC components. Then, a two-dimensional fast Fourier transform is performed on the clutter-filtered radar echo signal along both the range and velocity dimensions to construct a range-Doppler matrix, thus achieving initial target saliency enhancement.

[0046] In this embodiment, the distance dimension resolution is determined by the equivalent bandwidth. Let... For linear frequency modulation slope, Number of sampling points per chirp For ADC sampling rate, =3×10⁸ m / s is the speed of light, then the equivalent bandwidth is... and distance resolution It can be represented as:

[0047] ; The maximum unambiguous velocity and chirp duration corresponding to the velocity dimension resolution In common configurations, this can be represented as: ; in, λ For radar operating wavelength, The duration of a single chirp. By performing a peak search on the range-Doppler matrix, the range and velocity parameters corresponding to the potential target can be obtained, providing the basic input for subsequent angle estimation and point cloud separation.

[0048] High-precision angle estimation based on the Capon algorithm: For each target peak detected in step one, array snapshot data within the corresponding range gate is extracted to construct the received signal covariance matrix. Then, Capon minimum variance spectrum estimation is performed by combining diagonal loading parameters, angle search range, and angle resolution to obtain high-precision azimuth and / or elevation angle information of the target. Range, velocity, and angle information are fused to construct a spatiotemporal dataset. Furthermore, the robustness of the covariance matrix is ​​improved through diagonal loading technology, and a global peak search is performed on the Capon angle spectrum to calculate the target's azimuth information. When the array structure supports elevation dimension measurement, the target's elevation angle information can also be acquired simultaneously.

[0049] Specifically, for each target energy peak identified in the range-Doppler domain, the corresponding range gate cell is first located, and the array snapshot data matrix of the radar received signal within that range gate is extracted. Let the number of snapshots be... , For the snapshot data matrix, As a diagonal loading factor, If the identity matrix is ​​used, then the received signal covariance matrix can be expressed as:

[0050] ; By introducing a diagonal loading technique into the covariance matrix, robust optimization of the covariance matrix can be performed, which can alleviate the problem of estimation instability under finite snapshot conditions and reduce the interference of false spectral peaks on the angle estimation results.

[0051] Construct a guide vector that matches the array structure The spatial power spectral density function is obtained based on the Capon minimum variance criterion: ; By performing a global peak search on the aforementioned angle spectrum, high-precision azimuth information of the target can be obtained. When the radar array supports elevation dimension measurement, the target's elevation angle information can also be obtained simultaneously. Subsequently, the calculated angle information is fused with the obtained range, velocity, and frame sequence information to form a spatiotemporal dataset.

[0052] In some application scenarios where further analysis of the target's micro-motion characteristics is required, after angle estimation is completed, the phase sequence or the target's local energy sequence can be combined to extract characteristic parameters such as the target's vibration amplitude and vibration frequency through methods such as least squares signal separation, so as to enhance the ability to identify different target motion patterns.

[0053] Potential target extraction based on CFAR detection: Based on a spatiotemporal dataset, a constant false alarm rate (CFAR) detection method is adopted. The local noise power is estimated in a sliding window manner, and the detection threshold is adaptively calculated by combining the number of training units, the number of guard units, and the false alarm rate parameter. Spatiotemporal units with signal strength higher than the detection threshold are marked as potential target points.

[0054] Specifically, using a spatiotemporal dataset as input, constant false alarm rate detection parameters are set, including at least the number of training units, the number of protection units, the detection window size, the target false alarm rate, and the threshold scaling factor.

[0055] A sliding window approach is used to traverse each spatiotemporal cell, estimating the local noise power in the training region surrounding the cell to be detected, while simultaneously eliminating the influence of the guard cell on the target main lobe energy. A detection threshold is adaptively calculated based on the local noise level, and spatiotemporal cells with signal strengths exceeding the threshold are marked as valid potential target points. For consecutive adjacent detection units belonging to the same target response, the detection unit is the spatiotemporal unit processed one by one during the traversal. When its signal strength exceeds the adaptive detection threshold, it is marked as a potential / effective potential target point. "Detection unit" is the term used during processing, and "spatiotemporal unit" is its attribute definition. Meeting the judgment criteria makes it a potential target point. Merging can be performed according to adjacency or energy extreme values ​​to obtain a stable set of candidate target points, thus outputting noise-suppressed potential target points. The echo is first preprocessed to construct a spatiotemporal raster dataset, then CFAR is used to detect potential targets per spatiotemporal unit to extract them, and then adjacent detection units from the same source are clustered and merged. Finally, the selected target units are converted into point cloud data.

[0056] The detected potential target points are clustered using an improved centroid method, dividing them into several target clusters. Initial clustering parameters are set, including initial cluster centers, lateral distance thresholds, longitudinal distance thresholds, velocity thresholds, iteration number thresholds, density thresholds, and convergence conditions. A target point cloud density factor and distance weights are introduced to dynamically calculate the centroid position of each candidate cluster, transforming the target distance and orientation attributes into lateral and longitudinal distances in a natural coordinate system. Thresholds are set for the lateral, longitudinal, and velocity attributes. εᵣ x , εᵣᵧ , εᵥ .

[0057] The radar point cloud data is merged based on a lateral distance threshold. For points within this threshold range, a weighted average of the target's lateral distances is applied, using the following formula: ; in, p Represents the target point cloud i The amplitude (i.e., the target point cloud) i (corresponding radar echo signal strength) rx Indicates the lateral distance of the point cloud. N 1 This indicates the number of point clouds after merging based on a horizontal distance threshold. R mx Indicates the lateral distance of the target.

[0058] Next, based on the initially divided points N1, while maintaining the same horizontal distance, a weighted average is applied for further subdivision according to the vertical distance threshold εᵣᵧ, which can be expressed as: ; in, p This represents the amplitude of the point cloud. ry Represents the vertical distance of the point cloud. N 2 This indicates the number of point clouds after merging based on the vertical distance threshold. R m ᵧ Indicates the vertical distance of the target.

[0059] Once the target location is determined, several dividing points N2 can be obtained. Then, the target velocity is further subdivided based on the velocity threshold εᵥ. By performing a weighted average along the velocity dimension, a unique point can be obtained, which can be represented as: ; Where p represents the amplitude of the point cloud, v represents the velocity of the point cloud, N3 represents the number of point clouds after merging according to the velocity threshold, and V m Indicates the speed of the target.

[0060] Finally, the remaining N classes will be divided by a threshold, and the cluster center of each class will be [R]. mx ,R m ᵧ,V m The target point allocation results are iteratively optimized until the convergence condition is met. Clusters with fewer than a preset threshold of points or a density lower than a set threshold are identified as noise clusters and removed, thus obtaining the target point cloud in the multi-target point cloud cluster.

[0061] The output potential target points are transformed from polar coordinates to Cartesian coordinates to obtain features such as lateral distance, longitudinal distance, and radial velocity of the point cloud. To adapt to the situation of multiple targets distributed close together and uneven point cloud density, an improved centroid method is adopted. That is, the algorithm adaptively adjusts the clustering threshold by calculating the cross-entropy between two target point clouds and their respective target center points to solve the problem of false merging, thereby optimizing the performance of the centroid clustering algorithm to cluster and separate potential target points.

[0062] Before clustering begins, initial parameters are set, including horizontal distance threshold, vertical distance threshold, velocity dimension threshold, initial cluster center generation method, maximum number of iterations, centroid update convergence threshold, minimum number of points within a cluster, and density threshold. Subsequently, an adaptive parameter centroid method based on the fusion of cross-entropy and cluster density is used to perform preliminary clustering, obtaining the initial clusters and their corresponding cluster centers.

[0063] During the clustering iteration process, the cross-entropy index between each point cloud and its respective cluster center as well as the centers of neighboring clusters is calculated to evaluate the rationality of the point cloud's assignment; simultaneously, the local density and global density of each cluster are calculated, specifically: (1) Clustering results partitioning. The point cloud data is partitioned. n There are clusters, each containing a certain number of point sets. C k ={ p 1, p 2,⋯, p m},in k ∈[1, n ], m It is a cluster C k The number of points.

[0064] (2) Define distance thresholds. Define a neighborhood for each point cloud and set the following distance thresholds: lateral distance threshold. thᵣ x The lateral distance threshold thᵣᵧ limits the distance difference between point clouds along the x-axis; the longitudinal distance threshold thᵥ limits the distance difference between point clouds along the y-axis; and the absolute velocity threshold thᵥ limits the distance difference between point clouds along the absolute velocity v. (The last part, "any two points," appears to be an unrelated fragment and is omitted from the translation.) pᵢ =(x ᵢ , yᵢ , vᵢ )and pⱼ =( xⱼ , yⱼ , vⱼ The domain relationship is determined by the following conditions: | xᵢ - xⱼ |≤ thᵣ x , | yᵢ−yⱼ |≤ thᵣᵧ , | vᵢ−vⱼ |≤ thᵥ Points that satisfy the above conditions pⱼ Considered to be at point pᵢ Within its neighborhood.

[0065] (3) Calculate the local density. For each cluster C k Calculate the local density of each point within the cluster. Specifically: for the cluster... C k Each point in pᵢ Count the number of points within its neighborhood. Nᵢ The formula for calculating the local density at this point is as follows: Dᵢ=Nᵢ / m ,in m For clusters C k The number of points.

[0066] (4) Calculate the global density. For cluster C k Local density of all points within Dᵢ Summing yields the overall density of the cluster. D k Then sum the overall densities of all clusters and divide by the total number of clusters. n To obtain the global density D The objective is to minimize the combined index of cross-entropy and global density, as shown in the following formulas:

[0067] ; ; In the formula, For cross-entropy, It is a probability matrix. This is the true probability matrix. Let c1 and c2 be the probability that the cluster centers of the two clusters closest to xi after clustering belong to the same cluster. Let c1 and c2 be the probability that the cluster centers of the two clusters closest to xi after clustering belong to the same cluster.

[0068] By combining global density D Cross-entropy of clustering results When the clustering algorithm searches for the minimum cross-entropy of the point cloud, the clusters will change accordingly, thus affecting the global density D of the clusters. When the global density is low, the algorithm will automatically adjust the clustering threshold and tend to maintain a larger clustering threshold to merge adjacent clusters. Conversely, when the global density is high, the algorithm tends to maintain a smaller clustering threshold to reduce erroneous merging caused by sparse point clouds.

[0069] Centroid clustering considers any two points as belonging to the same cluster if there is a path connecting them such that the distance between any two adjacent points does not exceed a threshold. This can lead to points far from the cluster center being excluded from the cluster center attribution threshold, causing the same target cluster to split into multiple target clusters during cross-entropy calculation. Adaptively adjusting the cluster center attribution threshold ensures that even points far from the cluster center remain within the threshold, effectively reducing the splitting of target clusters.

[0070] After the threshold is updated, the weighted distance from each potential target point to the centroid of each cluster is recalculated, and the target points are assigned to the cluster with the optimal distance. Then, the cluster centroid positions and point cloud assignment results are updated iteratively. For clusters with fewer than a preset threshold of points, density lower than a set threshold, or insufficient stability in continuous iterations, they are identified as noise clusters and removed, thus obtaining clean and independent multi-target point cloud clusters.

[0071] Based on target point clouds in a multi-target point cloud cluster and combined with estimated high-precision angle information, an iterative nearest-point algorithm is used to register point clouds in adjacent frames. The rotation and translation matrices are solved by minimizing the registration error to achieve inter-frame target point cloud matching. On this basis, a multi-unscented Kalman filter tracker is constructed based on the SORT framework. The target state vector is constructed using the target cluster centroid coordinates, Doppler velocity, and DOA angle. Target state prediction and updating are performed by incorporating process noise, observation noise, and initial covariance parameters. Specifically, the high-precision azimuth and elevation angles can be used as angle constraint weights in the registration process to correct the distance metric during point pair establishment, thus suppressing mismatches under range aliasing conditions.

[0072] Using the target point cloud in the current frame's multi-target point cloud cluster and the reference point cloud processed in the previous frame as input, the geometric center of the target cluster in the current frame is selected as the initial position for iteration, and the parameters of the iterative nearest point algorithm are set. The parameters include at least the maximum number of iterations, the nearest neighbor search radius, the registration error threshold, and the initial rotation matrix and translation matrix.

[0073] During registration, firstly, the nearest distance from the centroid of each target cluster in the current point cloud to the centroid of each cluster in the reference point cloud is calculated based on Euclidean distance, establishing a set of point pairs. Then, high-precision azimuth and pitch angles are introduced as angular constraint weights to correct the distance metric function. The nearest neighbors of the centroids of the current and reference point cloud clusters are calculated using Euclidean distance to construct an initial set of matched point pairs. High-precision azimuth and pitch angles are used as constraint weights to perform a weighted correction on the original Euclidean distance metric function to suppress mismatches in distance aliasing scenarios. Finally, the attitude error objective function is constructed, as shown in the following formula:

[0074] ; in, The centroid of the target point cloud cluster in the current frame; For the centroids of the corresponding target point cloud clusters in adjacent frames, The angle constraint weights are calculated from the azimuth and elevation angles.

[0075] The optimal rotation and translation matrices that minimize the error are obtained by solving singular value decomposition to achieve accurate matching of target point clouds in adjacent frames.

[0076] Based on point cloud registration, a multi-unscented Kalman filter is constructed using the simplified online real-time tracking SORT framework. The target state vector includes at least the three-dimensional coordinates of the cluster centroid, the projection components of the Doppler velocity in three-dimensional space, and azimuth and elevation angle information. Combining process noise covariance, observation noise covariance, initial error covariance, and Sigma point extension parameters, four types of parameters jointly participate in the covariance matrix prediction and update calculation, forming all the input parameters for constructing the state covariance matrix. The current frame target state and its covariance matrix can be predicted using a constant velocity model, which includes the following four steps:

[0077] 1. State input: Complete target state vector from the previous frame (3D coordinates of cluster centroid, 3D Doppler velocity projection, azimuth angle, pitch angle); 2. Noise input: Process noise covariance matrix; 3. Model parameters: Constant velocity state transition matrix; 4. Auxiliary parameters: Sigma point extension parameters, used for unscented transform sampling prediction; 5. Output: Predicted target state vector in the current frame + predicted state covariance matrix The actual detected values, namely the three-dimensional spatial coordinates of the target cluster centroid, the three-dimensional Doppler velocity projection components, and the high-precision azimuth and pitch angles, are used to form the UKF-filtered measurement vector Z, which is a new target state estimate, and thus outputs the smoothed multi-target trajectory points.

[0078] The Hungarian algorithm is employed to associate and match multi-target trajectory points with the detected target in the current frame, linking target point clouds in consecutive frames to the same motion trajectory. For targets that are briefly lost or falsely detected, a memory mechanism is implemented to record the target's state information before loss, and association thresholds, hit count thresholds, and loss count thresholds are set. When no matching point cloud is detected again within the preset threshold range, the corresponding trajectory is terminated and the corresponding trajectory data is deleted. Furthermore, the target's state position predicted by the filter at the next moment is used as the tracking threshold center. A joint cost matrix is ​​constructed by combining position deviation, velocity consistency, and angle consistency terms, and the Hungarian algorithm is used to solve for the globally optimal matching pair to improve the trajectory association accuracy in complex scenes.

[0079] Using the target state position predicted by the above filter in the next frame as the tracking threshold center, the Mahalanobis distance between it and the center of each target point cloud cluster detected in the current frame is calculated. A joint cost matrix is ​​constructed by combining the position deviation term, velocity consistency term, and angle consistency term. The joint cost matrix is ​​obtained by fusing azimuth and pitch angle correction distance metrics. The formula for the joint cost matrix is ​​as follows: ; in, ; ; ; ; ; In the formula, The cost matrix based on Euclidean distance and Mahalanobis distance is the first... i Line 1 j Column elements; For Euclidean distance function and Mahalanobis distance function, and The current frame number i The state vector of the tracker and the current frame's state vector. j The observation vector of each observation point trace. The observed covariance matrix is ​​used to predict the state covariance. For the observation matrix, For the updated state estimation error covariance matrix at time k+1, The covariance matrix of the observed noise, s e To observe the noise figure, The observed variances are the radial distance and radial velocity.

[0080] Simultaneously, an association threshold parameter is set. The joint cost matrix is ​​input into the Hungarian algorithm to solve for the globally optimal matching pair between the centroid observation information of all target clusters extracted after point cloud registration processing of the radar point cloud in the current frame and the multi-target tracks remaining after smoothing and association matching of the previous frame using unscented Kalman filtering. If a predicted trajectory is the theoretical trajectory of the target in the current frame that the tracker pre-calculates based on historical motion patterns, and there is no matching detection point cloud in the current frame, or if a detection point cloud cannot match any existing trajectory, it is marked as an unmatched item.

[0081] For successfully matched trajectory-detection pairs, the hit counter is updated. Consecutive hits refer to the same newly generated trajectory successfully matching with the measured target detection points in the current frame across multiple consecutive frames. New trajectories with consecutive hit counts exceeding the confirmation threshold are marked as confirmed trajectories and output. These new trajectories originate from detection point clouds that cannot match existing historical tracks in the current frame, representing newly emerging targets. These new trajectories are not predicted trajectories from historical frames; they are generated in the current frame based on new detection observations and become valid trajectories after verification through multiple consecutive hits. For unmatched confirmed trajectories, a memory mechanism is activated, marking them as pending confirmation and incrementing the loss counter. If a trajectory fails to rematch within the set loss counter threshold, or if the predicted position exceeds the radar's field of view, the target is determined to have left the scene, the corresponding trajectory is terminated, and the relevant trajectory data is deleted.

[0082] The above method can realize a complete processing flow from raw echo to multi-target trajectory output, and maintain good trajectory stability and continuity in scenarios with multiple target intersections, short-term occlusion and local false detection.

[0083] Example 2: In another embodiment, the constant false alarm rate (CFAR) detection method can select unit average type, maximum mean type, minimum mean type or sequential statistical type CFAR according to the characteristics of the scene; the initial cluster center can also be generated based on historical trajectory prediction results, point cloud peak positions or gridded statistical results to further improve the clustering stability in complex scenes.

[0084] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this invention.

[0085] The present invention also provides a computer-readable storage medium storing a computer program that can be used to execute the above-described... Figure 1 A radar multi-target point cloud registration method is provided.

[0086] The present invention also provides Figure 5The schematic diagram of the computer device shown is as follows: Figure 5 As shown, at the hardware level, this computer device includes a processor, internal bus, network interface, memory, and non-volatile memory, and may also include other hardware required for business operations. The processor reads the corresponding computer program from the non-volatile memory into memory and then executes it to achieve the above. Figure 1 A radar multi-target point cloud registration method is provided.

[0087] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

Claims

1. A radar multi-target point cloud registration method, characterized in that, include: Radar echo signals from multiple moving targets are acquired frame by frame; Based on the radar echo signal, a spatiotemporal point cloud signal for each frame is constructed; The CFAR detection method is used to identify the region of spatiotemporal point cloud signal with a signal strength higher than a preset signal strength threshold in each frame as the potential target point cloud region for each frame. The potential target point cloud region of each frame is divided into several target point cloud clusters, and initial clustering parameters are set to perform clustering processing on the target point clouds in each target point cloud cluster. During the iterative process of clustering processing, the radar echo signal intensity corresponding to the target point cloud is introduced as a range weight to determine the centroid position of each target point cloud cluster. Based on the centroid position of the target point cloud cluster, a comprehensive evaluation index is constructed to assess the rationality of point cloud assignment. With the goal of minimizing the comprehensive evaluation index, the initial clustering parameters are adaptively adjusted to obtain the point cloud redistribution result in each target point cloud cluster. Based on the point cloud redistribution result in each target point cloud cluster, multi-target point cloud clusters in the potential target point cloud region of each frame are determined. The iterative nearest point algorithm is used to register the target point clouds in the multi-target point cloud clusters in the potential target point cloud region of adjacent frames, so as to realize the multi-target point cloud registration of the radar.

2. The radar multi-target point cloud registration method as described in claim 1, characterized in that, Based on the radar echo signal, a spatiotemporal point cloud signal is constructed, specifically including: Two-dimensional fast Fourier transforms are performed on the radar echo signal in the range and velocity dimensions respectively to obtain the range-Doppler matrix; range and velocity parameters of potential target echoes are extracted from the range-Doppler matrix by peak search; For the peak values ​​of the range and velocity parameters of potential target echoes, array snapshot data corresponding to the potential target echoes within the range gate is extracted. Based on the array snapshot data, a covariance matrix of the echo received signal is constructed. A diagonal loading technique is introduced into the covariance matrix of the echo received signal to robustly optimize the covariance matrix. According to the robustly optimized covariance matrix, a steering vector matching the array snapshot data structure is constructed. Based on the steering vector and the robustly optimized covariance matrix, the spatial power spectral density function is determined using the Capon minimum variance criterion. A global peak search is performed on the spatial power spectral density function to obtain the azimuth information of the potential target echo. The range parameters, velocity parameters, azimuth information, and frame sequence information of the potential target echo are fused to obtain a spatiotemporal point cloud signal.

3. The radar multi-target point cloud registration method as described in claim 1, characterized in that, The initial clustering parameters include: initial cluster centers, lateral distance threshold, longitudinal distance threshold, velocity threshold, iteration number threshold, density threshold, and convergence condition.

4. The radar multi-target point cloud registration method as described in claim 3, characterized in that, The introduction of target point cloud distance weights to determine the centroid position of each target point cloud cluster specifically includes: Transform the target point clouds in each target point cloud cluster from the polar coordinate system to the rectangular coordinate system to obtain the lateral distance, longitudinal distance and velocity of each target point cloud; The target point clouds whose lateral distance is within the lateral distance threshold are merged to obtain a first target point cloud set; the lateral distance of each target point cloud in the first target point cloud set is weighted by the point cloud amplitude of each target point cloud in the first target point cloud set, and the lateral distance of each target point cloud in the first target point cloud set is weighted to obtain the target point cloud lateral distance of each target point cloud cluster. Merge the target point clouds in the first target point cloud set that are within the longitudinal distance threshold to obtain the second target point cloud set; use the point cloud amplitude of each target point cloud in the second target point cloud set as the longitudinal distance weight, and perform a weighted average of the longitudinal distances of each target point cloud in the second target point cloud cluster set to obtain the longitudinal distance of the target point cloud in each target point cloud cluster. The target point clouds within the velocity threshold in the second target point cloud set are merged to obtain the third target point cloud set; the velocity of each target point cloud in the third target point cloud set is weighted by the point cloud amplitude of each target point cloud in the third target point cloud set, and the velocity of the target point cloud in each target point cloud cluster is obtained by weighted averaging. The horizontal distance and vertical distance of the target point cloud are determined as the centroid positions of the target point cloud cluster, and each centroid position corresponds to the velocity of the target point cloud cluster.

5. The radar multi-target point cloud registration method as described in claim 1, characterized in that, The comprehensive evaluation index for assessing the rationality of point cloud attribution, constructed based on the centroid location of the target point cloud region, specifically includes: For each target point cloud in a target point cloud cluster, determine the centroid position of the target point cloud and its own target point cloud cluster, as well as the cross-entropy between the target point cloud and the centroid positions of its neighboring target point cloud clusters; The global density of each target point cloud cluster is determined based on the local density of each target point cloud in each target point cloud cluster. Based on the cross-entropy and the global density of the target point cloud cluster, a comprehensive evaluation index is constructed to assess the rationality of point cloud attribution.

6. The radar multi-target point cloud registration method as described in claim 5, characterized in that, The local density of the target point cloud in the target point cloud cluster is determined by the ratio of the number of point clouds in the target point cloud within a preset range to the total number of target point clouds in the target point cloud cluster; the global density is determined by summing the local densities of each target point cloud in each target point cloud cluster and then dividing by the total number of target point cloud clusters.

7. The radar multi-target point cloud registration method as described in claim 2, characterized in that, The step of using the iterative nearest-neighbor algorithm to register target point clouds in multi-target point cloud clusters within potential target point cloud regions of adjacent frames specifically includes: The geometric center of the multi-target point cloud cluster in the potential target point cloud region of the current frame is selected as the registration point set, and the geometric center of the multi-target point cloud cluster after processing in the previous frame is selected as the reference point set. The nearest distance between each point in the set of points to be registered and each point in the set of reference points is calculated based on Euclidean distance, and an initial set of point pairs is established. The azimuth information in the spatiotemporal point cloud signal is used as the angle constraint weight to correct the nearest distance in the initial set of point pairs, and a corrected set of point pairs is obtained. Based on the centroid positions of the multi-target point cloud clusters in the potential target point cloud regions of the current frame and the previous frame, as well as the azimuth information in the spatiotemporal point cloud signal, an attitude error objective function is constructed. By performing singular value decomposition on the set of correction point pairs, the optimal rotation and translation matrices that minimize the attitude error objective function are solved. The optimal rotation and translation matrices are then used to perform coordinate transformation on the target point clouds in the multi-target point cloud clusters of the current frame, thereby realizing the target point clouds in the multi-target point cloud clusters in the potential target point cloud regions of adjacent frames.

8. The radar multi-target point cloud registration method as described in claim 1, characterized in that, After achieving registration of target point clouds in multi-target point cloud clusters of adjacent frames, the following steps are also included: A multi-unscented Kalman filter tracker is constructed based on the SORT framework. The target point cloud state vector is formed by the centroid position coordinates, the projection components of the Doppler velocity in three-dimensional space, and the azimuth and pitch information. Using the target point cloud state vector, state covariance matrix, and Sigma point extension parameters as input parameters for the constant velocity model, the system outputs the target state value and its covariance matrix of the target point cloud in the current frame. Based on the actual detected point cloud state value as the observation value, the system performs unscented Kalman filtering to update the target point cloud state estimate and outputs the smooth trajectory points of the target point cloud in the current frame.

9. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the method described in any one of claims 1 to 8.

10. A computer device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described in any one of claims 1 to 8.

Images