A millimeter-wave radar point cloud quality sensing tracking method based on adaptive filtering and multi-feature hierarchical clustering
By employing adaptive filtering and multi-feature hierarchical clustering, the problems of fixed filtering parameters and clustering dependency in millimeter-wave radar point cloud processing are solved, improving clustering accuracy and tracking stability, and achieving higher reliability in environmental perception.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ZHONGYUAN ENGINEERING COLLEGE
- Filing Date
- 2026-04-30
- Publication Date
- 2026-06-30
Smart Images
Figure CN122307500A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of autonomous driving environmental perception technology, and in particular to a millimeter-wave radar point cloud quality perception and tracking method based on adaptive filtering and multi-feature hierarchical clustering. Background Technology
[0002] In the field of point cloud processing for autonomous driving millimeter-wave radar, current research mainly focuses on two major tasks: point cloud clustering and multi-target tracking. Regarding point cloud clustering, density-based methods such as DBSCAN can handle noisy and non-convex targets to some extent, but their performance is heavily reliant on manually set parameters such as neighborhood radius and minimum number of points, making it difficult to adapt to the inherent sparsity, uneven distribution, and noise interference of millimeter-wave radar point clouds. Furthermore, commonly used Euclidean distance metrics mostly consider only spatial location information, failing to fully integrate the correlation between point cloud characteristics such as velocity and RCS (radar cross section), which limits their clustering accuracy and robustness in real-world scenarios. In multi-target tracking, tracking frameworks represented by Kalman filtering and its nonlinear extensions have been widely applied. While improved methods such as interactive multi-model approaches can partially address target maneuvering behavior, they generally lack mechanisms for perceiving and utilizing the quality of clustering results. Since existing methods often struggle to dynamically adjust motion model weights or data association thresholds based on the reliability of cluster outputs, the performance of tracking systems is prone to degradation when point cloud quality fluctuates, targets are occluded, or there is strong noise interference. There is still room for improvement in the ability to maintain stable tracking in real-world scenarios. Summary of the Invention
[0003] This invention addresses the problems existing in current millimeter-wave radar point cloud processing methods, such as fixed filtering parameters, strong reliance on human experience in the clustering process, and insufficient output quality perception in the tracking stage. It proposes a millimeter-wave radar point cloud quality perception tracking method based on adaptive filtering and multi-feature hierarchical clustering, which can improve the reliability of environmental perception of autonomous driving systems in complex scenarios.
[0004] To achieve the above objectives, the present invention adopts the following technical solution:
[0005] This invention proposes a millimeter-wave radar point cloud quality sensing and tracking method based on adaptive filtering and multi-feature hierarchical clustering, comprising:
[0006] Obtain the original point cloud set output by the millimeter-wave radar, perform adaptive cascade filtering on the original point cloud set, and output the effective point cloud set;
[0007] The effective point cloud set is subjected to multi-feature hierarchical clustering, including coarse clustering based on multi-feature fusion, calculating the coarse clustering effectiveness score of each coarse cluster and selecting the retained clusters accordingly, and performing fine clustering on the retained clusters to generate the observation target set.
[0008] The set of observed targets is associated with the set of existing tracking trajectories, a comprehensive quality score for each observed target is calculated, a quality confidence score is obtained by mapping the comprehensive quality score, an adaptive association threshold is calculated based on the quality confidence score, and the matching of observations and trajectories is completed based on the adaptive association threshold.
[0009] The state of the successfully matched trajectory is updated based on the motion model, and the tracking result is output.
[0010] Furthermore, the adaptive cascaded filtering includes:
[0011] Based on the preset effective detection area, select point clouds within the effective spatial range from the original point cloud set;
[0012] Calculate the mean and standard deviation of the RCS features of the point cloud within the effective spatial range, set the RCS threshold according to the overall reflection intensity level and dispersion of the point cloud in the current frame, and divide the point cloud into a high RCS point set and a low RCS point set.
[0013] Density-based filtering is applied to the high RCS point set to remove isolated and random noise points, resulting in a high RCS effective point set.
[0014] The set of low RCS points is merged with the set of high RCS valid points to obtain the set of valid point clouds.
[0015] Furthermore, the multi-feature hierarchical clustering includes:
[0016] Based on spatial features, velocity features, and RCS features, calculate the composite distance metric between any two point clouds in the effective point cloud set;
[0017] Based on the aforementioned composite distance metric, a density clustering algorithm is used to coarsely cluster all valid point clouds to obtain preliminary coarse clusters.
[0018] For each coarse cluster, calculate its velocity consistency score and spatial compactness score, and then weight and fuse them to obtain the coarse cluster effectiveness score;
[0019] Retain coarse clusters whose coarse clustering effectiveness score reaches a preset threshold, and discard coarse clusters that do not reach the threshold;
[0020] For the retained coarse clusters, fine clustering is performed based on spatial characteristics to obtain the fine cluster observation targets.
[0021] Furthermore, the calculation of the overall quality score for each observed target includes:
[0022] For each fine cluster observation target, the scores of its four dimensions—point density score, velocity consistency score, RCS quality score, and spatial compactness score—are combined and weighted to obtain a comprehensive quality score.
[0023] Furthermore, the point density score is calculated based on the ratio of the number of points in a cluster to the spatial area occupied by the cluster, and takes the smaller value between 1 and 1.
[0024] The RCS quality score is calculated based on the ratio of the mean to the standard deviation of the RCS values in the cluster, and takes the smaller value between 1 and 1.
[0025] Furthermore, the quality confidence level is calculated based on the following method:
[0026] The overall quality score is converted into a quality confidence level between 0 and 1 through a nonlinear mapping function, so that the higher the overall quality score, the closer the quality confidence level is to 1.
[0027] Furthermore, the adaptive association threshold is determined based on the following method:
[0028]
[0029] in Representing the trajectory With predicted trajectory Adaptive correlation threshold between them For the preset threshold coefficient, For trajectory Optimal motion model at the current moment The new information covariance matrix, For quality confidence level.
[0030] Furthermore, associating the observed target set with the existing tracking trajectory set includes:
[0031] Calculate the Mahalanobis distance between the observed and predicted trajectories. When the Mahalanobis distance is less than or equal to the adaptive correlation threshold, the correlation cost is determined by the Mahalanobis distance and the quality confidence score. When the Mahalanobis distance is greater than the adaptive correlation threshold, the correlation cost is set to infinity.
[0032] Furthermore, the state update of the successfully matched trajectory based on the motion model, and the output tracking results include:
[0033] Update the posterior probability of each motion model, wherein the quality confidence is incorporated into the model probability update process;
[0034] The motion model with the highest posterior probability is selected as the optimal motion model at the current moment. The Kalman gain is calculated based on the optimal motion model, and the trajectory state estimate is updated.
[0035] The motion model includes at least two of the following models: uniform velocity model, uniform acceleration model, and constant rotation rate and velocity model.
[0036] The present invention also proposes a computer-readable storage medium storing a computer program thereon, wherein the computer program, when executed by a processor, implements the millimeter-wave radar point cloud quality perception and tracking method based on adaptive filtering and multi-feature hierarchical clustering as described above.
[0037] Compared with the prior art, the present invention has the following beneficial effects:
[0038] To address the problems in existing millimeter-wave radar point cloud processing methods, such as fixed filtering parameters, heavy reliance on human experience in clustering, and insufficient output quality perception in the tracking stage, this invention makes corresponding improvements. In the filtering stage, an adaptive cascaded filtering strategy based on the statistical characteristics and spatial distribution density of the radar cross section is adopted, achieving the distinction between noise and effective targets. This improves performance by approximately 12.7% compared to traditional fixed threshold methods, enhancing the retention of weakly reflective targets while suppressing noise. In the clustering stage, a hierarchical clustering algorithm with multi-feature fusion is proposed. This algorithm integrates spatial location, radial velocity, and RCS information to construct a composite distance metric and introduces a progressive processing flow of "coarse clustering - quality assessment - fine clustering," reducing the Davidson-Bolding index to 0.117, which is superior to DBSCAN (0.173) and K-means (0.343), helping to improve the performance of traditional methods in terms of feature utilization and parameter adaptability. In the tracking stage, the clustering quality assessment results are mapped to confidence weights and embedded into an interactive multi-model tracking framework, supporting dynamic adjustment of motion model probability updates and data association thresholds. Experimental results show that this method can reduce the root mean square error of the lateral position of the target trajectory by about 0.1052 meters to 0.1270 meters, improving the stability of the tracking system under point cloud quality fluctuations and real-world scenarios. Attached Figure Description
[0039] Figure 1 This is one of the flowcharts of a millimeter-wave radar point cloud quality perception and tracking method based on adaptive filtering and multi-feature hierarchical clustering according to an embodiment of the present invention.
[0040] Figure 2 This is the second flowchart of an embodiment of the present invention, illustrating a millimeter-wave radar point cloud quality perception and tracking method based on adaptive filtering and multi-feature hierarchical clustering.
[0041] Figure 3 This is a schematic diagram of an experimental scenario provided in an embodiment of the present invention;
[0042] Figure 4 Example diagram of raw point cloud data from millimeter-wave radar provided in an embodiment of the present invention;
[0043] Figure 5 A comparison chart of filtering effects provided in an embodiment of the present invention;
[0044] Figure 6 This is a comparison chart of clustering effects provided in an embodiment of the present invention;
[0045] Figure 7 A trajectory tracking comparison diagram provided for an embodiment of the present invention. Detailed Implementation
[0046] The present invention will be further explained below with reference to the accompanying drawings and specific embodiments:
[0047] like Figure 1 As shown, a millimeter-wave radar point cloud quality perception tracking method based on adaptive filtering and multi-feature hierarchical clustering includes:
[0048] Obtain the original point cloud set output by the millimeter-wave radar, perform adaptive cascade filtering on the original point cloud set, and output the effective point cloud set;
[0049] The effective point cloud set is subjected to multi-feature hierarchical clustering, including coarse clustering based on multi-feature fusion, calculating the coarse clustering effectiveness score of each coarse cluster and selecting the retained clusters accordingly, and performing fine clustering on the retained clusters to generate the observation target set.
[0050] The set of observed targets is associated with the set of existing tracking trajectories, a comprehensive quality score for each observed target is calculated, a quality confidence score is obtained by mapping the comprehensive quality score, an adaptive association threshold is calculated based on the quality confidence score, and the matching of observations and trajectories is completed based on the adaptive association threshold.
[0051] The state of the successfully matched trajectory is updated based on the motion model, and the tracking result is output.
[0052] Furthermore, the adaptive cascaded filtering includes:
[0053] Based on the preset effective detection area, select point clouds within the effective spatial range from the original point cloud set;
[0054] Calculate the mean and standard deviation of the RCS features of the point cloud within the effective spatial range, set the RCS threshold according to the overall reflection intensity level and dispersion of the point cloud in the current frame, and divide the point cloud into a high RCS point set and a low RCS point set.
[0055] Density-based filtering is applied to the high RCS point set to remove isolated and random noise points, resulting in a high RCS effective point set.
[0056] The set of low RCS points is merged with the set of high RCS valid points to obtain the set of valid point clouds.
[0057] Furthermore, the multi-feature hierarchical clustering includes:
[0058] Based on spatial features, velocity features, and RCS features, calculate the composite distance metric between any two point clouds in the effective point cloud set;
[0059] Based on the aforementioned composite distance metric, a density clustering algorithm is used to coarsely cluster all valid point clouds to obtain preliminary coarse clusters.
[0060] For each coarse cluster, calculate its velocity consistency score and spatial compactness score, and then weight and fuse them to obtain the coarse cluster effectiveness score;
[0061] Retain coarse clusters whose coarse clustering effectiveness score reaches a preset threshold, and discard coarse clusters that do not reach the threshold;
[0062] For the retained coarse clusters, fine clustering is performed based on spatial characteristics to obtain the fine cluster observation targets.
[0063] Furthermore, the calculation of the overall quality score for each observed target includes:
[0064] For each fine cluster observation target, the scores of its four dimensions—point density score, velocity consistency score, RCS quality score, and spatial compactness score—are combined and weighted to obtain a comprehensive quality score.
[0065] Furthermore, the point density score is calculated based on the ratio of the number of points in a cluster to the spatial area occupied by the cluster, and takes the smaller value between 1 and 1.
[0066] The RCS quality score is calculated based on the ratio of the mean to the standard deviation of the RCS values in the cluster, and takes the smaller value between 1 and 1.
[0067] Furthermore, the quality confidence level is calculated based on the following method:
[0068] The overall quality score is converted into a quality confidence level between 0 and 1 through a nonlinear mapping function, so that the higher the overall quality score, the closer the quality confidence level is to 1.
[0069] Furthermore, the adaptive association threshold is determined based on the following method:
[0070]
[0071] in Representing the trajectory With predicted trajectory Adaptive correlation threshold between them For the preset threshold coefficient, For trajectory Optimal motion model at the current moment The new information covariance matrix, For quality confidence level.
[0072] Furthermore, associating the observed target set with the existing tracking trajectory set includes:
[0073] Calculate the Mahalanobis distance between the observed and predicted trajectories. When the Mahalanobis distance is less than or equal to the adaptive correlation threshold, the correlation cost is determined by the Mahalanobis distance and the quality confidence score. When the Mahalanobis distance is greater than the adaptive correlation threshold, the correlation cost is set to infinity.
[0074] Furthermore, the state update of the successfully matched trajectory based on the motion model, and the output tracking results include:
[0075] Update the posterior probability of each motion model, wherein the quality confidence is incorporated into the model probability update process;
[0076] The motion model with the highest posterior probability is selected as the optimal motion model at the current moment. The Kalman gain is calculated based on the optimal motion model, and the trajectory state estimate is updated.
[0077] The motion model includes at least two of the following models: uniform velocity model, uniform acceleration model, and constant rotation rate and velocity model.
[0078] In the filtering stage, this invention employs an adaptive cascaded filtering strategy based on the statistical characteristics and spatial distribution density of radar cross sections, achieving the distinction between noise and effective targets, and enhancing the retention capability of weakly reflective targets while suppressing noise. In the clustering stage, a hierarchical clustering algorithm with multi-feature fusion is proposed, integrating spatial location, radial velocity, and RCS information to construct a composite distance metric, and introducing a progressive processing flow of "coarse clustering - quality assessment - fine clustering," which helps improve the performance of traditional methods in terms of feature utilization and parameter adaptability. In the tracking stage, the clustering quality assessment results are mapped to confidence weights and embedded into an interactive multi-model tracking framework, supporting dynamic adjustment of motion model probability updates and data association thresholds.
[0079] like Figure 2 As shown, another millimeter-wave radar point cloud quality sensing and tracking method based on adaptive filtering and multi-feature hierarchical clustering according to the present invention includes:
[0080] Step 1: Adaptive Cascade Filtering
[0081] 1. Spatial Range Constraints: Based on the preset effective detection area, the original point cloud dataset from the ARS 408-21 millimeter-wave radar is used. Select the point cloud set within the effective spatial range. .
[0082] 2. RCS Adaptive Filtering: Calculating the Point Cloud Set The mean of the RCS features of all points and standard deviation A threshold is set based on the overall reflection intensity level and dispersion of the current frame point cloud. Divide the point cloud into high RCS point sets and low RCS point set .
[0083] 3. Density filtering: for Density-based filtering is performed to remove isolated and random noise points, resulting in a high RCS effective point set. Output the final valid point cloud set. .
[0084]
[0085] Step 2: Multi-feature hierarchical clustering
[0086] 1. Composite distance metric:
[0087] for In the point cloud, the spatial weight between any two point clouds is calculated based on the Gaussian kernel function. Speed weight And integrate them into a comprehensive weight. Considering the product of spatial and velocity weights, a comprehensive weight is defined for each point cloud. The comprehensive weight of a given point cloud is summed with the comprehensive weights of all other point clouds, and the maximum value among these sums is taken as the feature value characterizing the core strength of that point cloud. The feature value reflecting the core strength of each point cloud is calculated. Point clouds with the top 80% of feature values are selected. The absolute differences between each pair of these point clouds in spatial, velocity, and RCS features are calculated and normalized. A composite distance metric is then formed by linearly weighted summation. .
[0088]
[0089]
[0090] 2. Clustering and Quality Assessment
[0091] The neighborhood radius of the DBSCAN algorithm is determined based on the composite distance percentile of all point cloud pairs. Coarse clustering was performed on all point clouds to obtain preliminary cluster divisions. The minimum number of neighborhood points required for the core points in the DBSCAN algorithm is determined proportionally to the total size of the point cloud. .
[0092]
[0093] For each coarse cluster Calculate its speed consistency score and space compactness score Weighted fusion yields the coarse clustering effectiveness score. By experimentally observing the actual situation of clusters corresponding to different scores, a threshold is set, and only clusters that meet the requirements of coarse clustering effectiveness score are retained.
[0094]
[0095] in, This represents the number of point clouds in the cluster; The speed of the point cloud.
[0096]
[0097]
[0098] in, ; This is the average of the coordinates of all points within the cluster; For point clouds The plane coordinates.
[0099]
[0100] For the retained high-quality coarse clusters, fine clustering is performed based on spatial features to separate multiple targets that may be stuck together.
[0101]
[0102] Step 3: Quality-Driven Multi-Model Tracking
[0103] 1. Observation and Trajectory Initialization: Transform the fine clustering results into a set of observation targets. and state vector Initialize or update the tracking trajectory set, and define three motion models. ,in Represents a uniform velocity model. Represents a uniform acceleration model. Representing a constant rotational speed and velocity model, using trajectory based on Kalman filtering. Specified motion model The state transition matrix is obtained The uncertainty of state estimation at time t and the uncertainty of predicted state estimation. At frame rate, the set of all trajectories the system is tracking is: A single trajectory serves as the input to the tracking algorithm, containing information such as state estimation, covariance, and historical information.
[0104]
[0105]
[0106] in, ;
[0107] 2. Quality confidence mapping:
[0108] For generating observation clusters The overall quality is evaluated, and the quality score is mapped to a confidence level. First, the cluster is calculated. The point density score is calculated; a higher score indicates a greater likelihood that the cluster is a target rather than sparse noise. The RCS quality score of the cluster is also calculated; a high cluster score indicates good target reflection characteristics. The scores from the four dimensions of density, velocity consistency, RCS quality, and spatial compactness are weighted and fused to obtain the cluster... Overall quality score Mapping the overall quality score to a quality confidence level. .
[0109]
[0110] in, For clusters Point density score, Indicates the size of the cluster (i.e., the cluster size). (number of point clouds included) Cluster Area in space The normalized density coefficient; For clusters RCS quality score, Cluster The mean of all point cloud RCS values, Cluster The standard deviation of all point cloud RCS values, To represent a very small positive number, This is the signal-to-noise ratio scaling factor; These are preset slope control parameters; This is the threshold for the overall quality score.
[0111] 3. Data association and status update:
[0112] Update the posterior probability of each motion model By incorporating quality confidence into the model probability update for tracking, quality-aware tracking is achieved. The motion model with the highest posterior probability is selected as the optimal motion model for the current trajectory. And calculate its new information covariance matrix. .
[0113]
[0114]
[0115] in, For observation In the model The likelihood function under the given conditions; .
[0116] The trajectory is calculated by combining the information covariance matrix and quality confidence score under the optimal model. With predicted trajectory Adaptive correlation thresholds between them. Based on the trajectory. With predicted trajectory Mahalanobis distance between and adaptive correlation threshold Calculate trajectory With predicted trajectory Related costs The global nearest neighbor association is used to match observations with trajectories.
[0117]
[0118]
[0119] in, , For the preset threshold coefficient, For trajectory Optimal motion model at the current moment The new information covariance matrix, For quality confidence level.
[0120] Calculate the Kalman gain, use the Kalman gain and associated observations to update the state estimate and its uncertainty, and prepare for processing in the next frame.
[0121] Furthermore, to verify the effectiveness of the present invention, the following experiments were conducted:
[0122] Experimental data acquisition scenarios such as Figure 3 As shown.
[0123] Based on the spatial constraints performed in step one to remove distant noise and irrelevant background point clouds, the point cloud set within the effective spatial range is... like Figure 4 As shown.
[0124] Adaptive cascaded filtering and fixed threshold filtering algorithms are applied to the original point cloud to remove noise and retain valid target points. The filtering effects are compared below. Figure 5 As shown in the figure. Judging from the filtering effects of the two methods, fixed threshold filtering has a weaker ability to distinguish noise and retains more interference points, while adaptive cascaded filtering can better identify and filter out noise, and retains more effective target points.
[0125] The point cloud after adaptive cascaded filtering in step two is clustered using the Multi-Feature Hierarchical Clustering (MFHC), DBSCAN, and K-means clustering algorithms, respectively. The clustering results are compared below. Figure 6 As shown, DBSCAN has many gray noise points due to density threshold limitations, and K-means, although it can divide regular clusters, has misclassification issues. The MFHC algorithm performs best, preserving the target structure while reducing noise interference, and its overall clustering effect is better than the former two.
[0126] Based on step three, we compare and analyze the experimental results of the Quality-Driven Multi-Model Tracking (QD-MMTA) algorithm and the Quality-Driven Single-Model Tracking (SMTA) algorithm. Figure 7 The image shows a trajectory comparison between two tracking algorithms. GPS-Veh-A and GPS-Veh-B represent the ground truth trajectories of the two vehicles located by GPS. QD-MMTA-Veh-A and QD-MMTA-Veh-B represent the trajectories of the two vehicles obtained using the QD-MMTA algorithm, and SMTA-A and SMTA-B represent the trajectories of the two vehicles obtained using the SMTA algorithm. The scatter points of the QD-MMTA-Veh-A and QD-MMTA-Veh-B trajectories are more concentrated, with less dispersion, and are closer to their corresponding ground truth trajectories.
[0127] Based on the above embodiments, the present invention also proposes a computer-readable storage medium storing a computer program thereon, wherein the computer program, when executed by a processor, implements the millimeter-wave radar point cloud quality perception and tracking method based on adaptive filtering and multi-feature hierarchical clustering as described above.
[0128] The above description is only a preferred embodiment of the present invention. It should be noted that those skilled in the art can make several improvements and modifications without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. A millimeter wave radar point cloud quality-aware tracking method based on adaptive filtering and multi-feature hierarchical clustering, characterized in that, include: Obtain the original point cloud set output by the millimeter-wave radar, perform adaptive cascade filtering on the original point cloud set, and output the effective point cloud set; The effective point cloud set is subjected to multi-feature hierarchical clustering, including coarse clustering based on multi-feature fusion, calculating the coarse clustering effectiveness score of each coarse cluster and selecting the retained clusters accordingly, and performing fine clustering on the retained clusters to generate the observation target set. The set of observed targets is associated with the set of existing tracking trajectories, a comprehensive quality score for each observed target is calculated, a quality confidence score is obtained by mapping the comprehensive quality score, an adaptive association threshold is calculated based on the quality confidence score, and the matching of observations and trajectories is completed based on the adaptive association threshold. The state of the successfully matched trajectory is updated based on the motion model, and the tracking result is output.
2. The method of claim 1, wherein, The adaptive cascaded filtering includes: Based on the preset effective detection area, select point clouds within the effective spatial range from the original point cloud set; Calculate the mean and standard deviation of the RCS features of the point cloud within the effective spatial range, set the RCS threshold according to the overall reflection intensity level and dispersion of the point cloud in the current frame, and divide the point cloud into a high RCS point set and a low RCS point set. Density-based filtering is applied to the high RCS point set to remove isolated and random noise points, resulting in a high RCS effective point set. The set of low RCS points is merged with the set of high RCS valid points to obtain the set of valid point clouds.
3. The method of claim 1, wherein, The multi-feature hierarchical clustering includes: Based on spatial features, velocity features, and RCS features, calculate the composite distance metric between any two point clouds in the effective point cloud set; Based on the aforementioned composite distance metric, a density clustering algorithm is used to coarsely cluster all valid point clouds to obtain preliminary coarse clusters. For each coarse cluster, calculate its velocity consistency score and spatial compactness score, and then weight and fuse them to obtain the coarse cluster effectiveness score; Retain coarse clusters whose coarse clustering effectiveness score reaches a preset threshold, and discard coarse clusters that do not reach the threshold; For the retained coarse clusters, fine clustering is performed based on spatial characteristics to obtain the fine cluster observation targets.
4. The method of claim 1, wherein, The calculation of the overall quality score for each observed target includes: For each fine cluster observation target, the scores of its four dimensions—point density score, velocity consistency score, RCS quality score, and spatial compactness score—are combined and weighted to obtain a comprehensive quality score.
5. The method of claim 4, wherein, The point density score is calculated based on the ratio of the number of points in a cluster to the spatial area occupied by the cluster, and takes the smaller value between 1 and 1. The RCS quality score is calculated based on the ratio of the mean to the standard deviation of the RCS values in the cluster, and takes the smaller value between 1 and 1.
6. The method of claim 1, wherein, The quality confidence level is calculated based on the following method: The overall quality score is converted into a quality confidence level between 0 and 1 through a nonlinear mapping function, so that the higher the overall quality score, the closer the quality confidence level is to 1.
7. The method of claim 1, wherein, The adaptive association threshold is determined based on the following method: wherein represents a trajectory between a predicted trajectory and an adaptive association threshold, is a preset threshold coefficient, is a trajectory optimal motion model of a new information covariance matrix at a current time, is a quality confidence.
8. The method of claim 1, wherein, Associating the observed target set with the existing tracking trajectory set includes: Calculate the Mahalanobis distance between the observed and predicted trajectories. When the Mahalanobis distance is less than or equal to the adaptive correlation threshold, the correlation cost is determined by the Mahalanobis distance and the quality confidence score. When the Mahalanobis distance is greater than the adaptive correlation threshold, the correlation cost is set to infinity.
9. The method of claim 1, wherein, The state update of the successfully matched trajectory based on the motion model outputs the tracking results, including: Update the posterior probability of each motion model, wherein the quality confidence is incorporated into the model probability update process; The motion model with the highest posterior probability is selected as the optimal motion model at the current moment. The Kalman gain is calculated based on the optimal motion model, and the trajectory state estimate is updated. The motion model includes at least two of the following models: uniform velocity model, uniform acceleration model, and constant rotation rate and velocity model.
10. A computer-readable storage medium having stored thereon a computer program, characterized in that, When the computer program is executed by the processor, it implements a millimeter-wave radar point cloud quality perception and tracking method based on adaptive filtering and multi-feature hierarchical clustering as described in any one of claims 1 to 9.