A millimeter wave radar point cloud dynamic adaptive anisotropic clustering method

Abstract

The application relates to a kind of millimeter wave radar point cloud dynamic self-adapting anisotropy clustering methods, comprising: introducing dynamic parameter adaptive mechanism on the basis of traditional DBSCAN algorithm, according to the radial distance of point cloud dynamic calculation neighborhood radius and minimum point number, realize parameter adaptation;The principal component analysis is carried out to local neighborhood, and the ellipsoidal anisotropy neighborhood is constructed along the main direction stretching, to fit the real geometry of target;Utilize radial velocity and radar scattering cross section to calculate multipath index and give attenuation weight, suppress multipath false point;Combining pre-processing and post-processing realizes multi-stage false alarm filtering;The application effectively solves the problem of far and near scene clustering performance fluctuation, long strip target fracture and multipath false point interference, significantly improves the precision rate and recall rate of millimeter wave radar target detection in complex scene, and is suitable for the environment perception of automatic driving and intelligent transportation system.

Description

Technical Field

[0001] This invention relates to the fields of radar signal processing and point cloud clustering technology, and in particular to a dynamic adaptive anisotropic clustering method for millimeter-wave radar point clouds. Background Technology

[0002] Millimeter-wave radar's high adaptability to harsh environments such as rain, fog, and lighting conditions makes it one of the core sensors for environmental perception in autonomous driving and intelligent transportation. However, millimeter-wave radar inherently faces the following technical challenges in the imaging process: The uneven distribution and density fluctuations of point clouds, with the number and density of points strongly influenced by the target distance and the relative pose of the radar and the target, result in denser points at close range and sparser points at distant range. This non-uniform distribution severely affects the stability of density-based clustering algorithms. Multipath interference occurs when radar signals are reflected multiple times and return, creating numerous "false points" around the real target. These false points typically have abnormal radial velocities or low radar cross sections, severely interfering with the detection of the target's true shape. False alarm interference, caused by sensor noise, environmental clutter, etc., usually has isolated distribution characteristics and low RCS value, which can easily lead to clustering algorithms generating false targets or misjudging.

[0003] DBSCAN (Density-Based Spatial Clustering of Applications with Noise), a classic algorithm in the field of density clustering, is widely used in point cloud processing. However, when dealing with millimeter-wave radar point clouds, traditional DBSCAN has two major drawbacks: Parameter rigidity: Using a globally uniform neighborhood radius and minimum number of points cannot adapt to the physical characteristics of the dynamic change of radar point cloud density with distance. Fixed parameters can easily lead to over-segmentation of nearby targets and missed detection of distant targets. Neighborhood model mismatch: Using an isotropic circular (or spherical) neighborhood search strategy makes it difficult to accurately characterize the geometric structure of typical targets such as vehicles with significant aspect ratios, which can easily lead to target fragmentation (undersegmentation) or multiple targets sticking together (oversegmentation). Summary of the Invention

[0004] In view of this, the present invention provides a dynamic adaptive anisotropic clustering method for millimeter-wave radar point clouds, which overcomes the shortcomings of the traditional DBSCAN algorithm in processing non-uniform and noisy point clouds of millimeter-wave radar. It effectively solves the problems of missed detection of distant targets, over-segmentation of nearby targets, fragmentation of long strip targets, and multipath pseudo-point interference, and significantly improves the accuracy of millimeter-wave radar target detection in complex scenarios.

[0005] To achieve the above objectives, this invention provides a dynamic adaptive anisotropic clustering method for millimeter-wave radar point clouds, comprising the following steps: S1. Data preprocessing and parameter initialization; Input a frame of millimeter-wave radar point cloud data, including each data point Spatial coordinates in the radar coordinate system radial velocity and radar cross section Calculate the radial distance of each data point. ; Regarding the radar cross section Below the preset threshold The data points are preprocessed to filter false alarms and marked as noise points, and then removed from the set of points to be clustered. S2. Calculate the dynamic adaptive parameters, including the dynamic neighborhood radius. and dynamic minimum number of points ; S3. Construct an ellipsoidal neighborhood based on principal component analysis: Traverse every unvisited data point. With the data points Principal component analysis is performed on the center and its local neighborhood to obtain the main direction of the point cloud distribution and to construct an ellipsoidal anisotropic neighborhood. S301, Local Principal Component Analysis: Based on the dynamic neighborhood radius... Define the initial neighborhood and obtain the point set within the neighborhood. And calculate the covariance matrix. ; ; in, This represents the geometric center of all points within the neighborhood, i.e., the cluster center. Indicates matrix transpose; For the covariance matrix Perform eigenvalue decomposition to obtain eigenvalues. , and the corresponding feature vectors , And satisfy ,in Used to indicate the main direction of point cloud distribution; S302. Define the length of the semi-axis of the ellipsoid, align the major semi-axis with the principal direction, and construct an anisotropic neighborhood of the ellipsoid. S303, Calculate candidate neighborhood points With the center point The offset vectors are respectively along the feature vectors Projection components of direction , Projection components of direction If satisfied Then determine the candidate neighborhood points. Belongs to the center point The ellipsoidal neighborhood; S4, Utilizing radial velocity and radar cross section The multipath evaluation function is used to calculate the multipath index and the point cloud weights. S5. The improved weighted DBSCAN algorithm is used to perform clustering region query and expansion on the dynamic adaptive parameters and point cloud weights to generate clusters; S6. Based on the consistency between the mean radar cross section and the velocity variance of the cluster, the cluster is filtered for false alarms, outliers of the cluster are removed and marked as final noise, and the point cloud set is obtained.

[0006] Preferably, the radial distance The expression is: .

[0007] Preferably, the dynamic neighborhood radius is calculated. The expression is: ; in, Indicates the preset basic neighborhood radius. The expansion factor represents the rate of increase of the control radius with distance. This indicates the preset maximum detection range. Represents the normalized distance factor. ; Calculate the dynamic minimum number of points The expression is: ; in, This represents the preset minimum number of neighborhood points. This represents the adjustment factor that controls the rate at which the minimum number of points increases with distance.

[0008] Preferably, the expressions for the semi-major and semi-minor axes of the ellipsoid are: ; ; in, Represents the semi-major axis of the ellipsoid. Represents the minor semi-axis of the ellipsoid. This represents the preset regularization parameter, used to control the sensitivity of the ellipsoid to anisotropy. This indicates the energy distribution ratio along the principal component direction.

[0009] Preferably, the multipath evaluation function calculation point Multipath index The expression is: ; in, Point radial velocity, Point The mean radial velocity of all points within the local neighborhood. Point The standard deviation of radial velocity at all points within a local neighborhood. Point radar cross section, This indicates a preset reference scattering cross section based on the target type. , All represent weighting coefficients; When the multipath index Greater than the preset threshold Judgment point Due to multipath interference, attenuation weights are introduced. The expression is: ; in, This represents the attenuation factor.

[0010] Preferably, the improved weighted DBSCAN algorithm performs clustering region query and expansion on the dynamic adaptive parameters and point cloud weights, specifically including the following steps: S501, Collection Point The expression for calculating the weighted effective number of points in the neighborhood point set is as follows: ; in, Point , neighborhood points; S502, if satisfied , will point Marked as core points and expanded into clusters, where Point The minimum number of points in the dynamic range; otherwise, the points... Mark it as a transient point.

[0011] Preferably, step S6 specifically includes the following steps: S601, For each cluster Perform internal consistency checks; S602, Calculate the mean within each cluster. and velocity variance ; S603, if the points in the cluster... satisfy Then the decision point Outliers are marked as final noise and removed. Point of Preset deviation threshold This represents the preset speed variance threshold; S604. Traverse all points in the cluster to obtain the final point cloud cluster set.

[0012] Compared with the prior art, the beneficial effects of the present invention are: This invention eliminates the limitation of traditional DBSCAN's globally fixed parameters in adapting to changes in point cloud density with distance by introducing a radial distance-driven dynamic parameter adaptive mechanism, achieving stable and consistent target clustering results across the entire range. The radial distance-driven parameter adaptive mechanism also eliminates the coupling effect of distance changes on clustering performance, solving the problem of drastic performance fluctuations in long-range and short-range scenes, and achieving stable target clustering results across the entire range. Furthermore, through PCA-based anisotropic ellipsoidal neighborhood modeling, the cluster boundaries can adaptively conform to the real spatial orientation of elongated targets such as vehicles, effectively solving the target breakage and adhesion problems caused by traditional circular neighborhoods. Simultaneously, by comprehensively utilizing a multipath interference exponential weighting strategy based on radial velocity and radar cross section, along with a multi-stage false alarm filtering mechanism combining preprocessing and post-processing, the interference from multipath false points and environmental clutter is significantly reduced without compromising the structural integrity of the real target, significantly improving the accuracy and recall rate of millimeter-wave radar target detection in complex scenes. Attached Figure Description

[0013] Figure 1 This is a schematic diagram of the clustering method of the present invention; Figure 2 This is a comparison chart of the clustering effects of the standard DBSCAN clustering algorithm and the improved dynamic DBSCAN clustering algorithm in an embodiment of the present invention. The left chart shows the clustering effect of the standard DBSCAN clustering algorithm, and the right chart shows the clustering effect of the improved dynamic DBSCAN clustering algorithm provided by the present invention. Detailed Implementation

[0014] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0015] This embodiment provides a dynamic adaptive anisotropic clustering method for millimeter-wave radar point clouds, involving millimeter-wave radar data processing and clustering techniques, and particularly an improved DBSCAN algorithm for processing data containing spatial coordinates. Target radial velocity With radar cross section Point cloud data of information; by introducing dynamic parameter adaptation, PCA-based anisotropic neighborhood modeling, multipath effect suppression, and false alarm point filtering, accurate clustering and identification of targets (such as vehicles) are achieved. Figure 1 As shown, the specific steps include: S1. Data preprocessing and parameter initialization; Input a frame of millimeter-wave radar point cloud data, including each data point Spatial coordinates in the radar coordinate system radial velocity and radar cross section , Here, the total number of point clouds in the current frame is represented, and the radial distance of each data point is calculated. The expression is: ; in, For point Two-dimensional plane coordinates in the radar coordinate system For point The radial distance, in meters (m); radar cross section Below the preset threshold The data points are preprocessed to filter false alarms and marked as noise points, and then removed from the set of points to be clustered.

[0016] S2, Pre-set the system's maximum detection distance Define the normalized distance factor. Calculate the dynamic adaptive parameters, including the dynamic neighborhood radius. and dynamic minimum number of points ; Preset basic neighborhood radius and expansion factor For distance is Calculate the dynamic neighborhood radius of the point. The expression is: ; in, Indicates the preset basic neighborhood radius. The expansion factor represents the rate of increase of the control radius with distance. This indicates the preset maximum detection range. Represents the normalized distance factor. This ensures that the neighborhood radius increases monotonically with increasing distance, in order to compensate for the decrease in density of distant points; Preset minimum number of neighborhood points and adjustment factor Calculate the dynamic minimum number of points The expression is: ; in, This represents the preset minimum number of neighborhood points. This represents the adjustment factor that controls the rate at which the minimum number of points increases with distance.

[0017] S3. Construct an ellipsoidal neighborhood based on principal component analysis: Traverse every unvisited data point. With the data points Principal component analysis is performed on the center and its local neighborhood to obtain the main direction of the point cloud distribution and to construct an ellipsoidal anisotropic neighborhood. S301, Local Principal Component Analysis: Based on the dynamic neighborhood radius Define the initial neighborhood and obtain the point set within the neighborhood. And calculate the covariance matrix. ; ; in, This represents the geometric center of all points within the neighborhood, i.e., the cluster center. Indicates matrix transpose; For covariance matrix Perform eigenvalue decomposition to obtain eigenvalues. , and the corresponding feature vectors , And satisfy ,in Used to indicate the main direction of point cloud distribution; S302, Criteria for constructing an ellipsoid: Define the length of the semi-axis of the ellipsoid, align the major semi-axis with the principal direction, and construct an anisotropic neighborhood of the ellipsoid. The expressions for the semi-major and semi-minor axes of the ellipsoid are: ; ; in, Represents the semi-major axis of the ellipsoid (unit: meter). Represents the minor axis of the ellipsoid (unit: meter). This represents the preset regularization parameter (dimensionless), used to control the sensitivity of the ellipsoid to anisotropy. This represents the energy distribution ratio along the principal component direction; a larger value indicates that the local point cloud distribution is more linear. S303, Calculate candidate neighborhood points With the center point The offset vectors are respectively along the feature vectors Projection components of direction , Projection components of direction If satisfied Then determine the candidate neighborhood points. Belongs to the center point The ellipsoidal neighborhood.

[0018] S4. To mitigate the impact of multipath pseudopoints on density connectivity, a multi-feature-based weighting factor is introduced: utilizing radial velocity. and radar cross section The multipath evaluation function is used to calculate the multipath index and the point cloud weights. Multipath evaluation function calculation points Multipath index The expression is: ; in, Point Radial velocity (unit: meters per second). Point Mean radial velocity of all points in the local neighborhood (unit: m / s). Point The standard deviation of radial velocity (in meters per second) at all points within a local neighborhood is used to measure the degree of velocity anomaly. Point Radar cross section (unit: dBsm). This indicates the preset reference scattering cross section (unit: dBsm) based on the target type (e.g., car, truck). , All represent weighting coefficients (dimensionless), used to balance velocity and The contribution of features; When multipath index Greater than the preset threshold Judgment point Due to multipath interference, attenuation weights are introduced. The expression is: ; in, The decay factor (dimensionless) is used to control the rate of weight decay; it is the multipath exponent of the true target point. The smaller the value, the higher the weight. The closer the multipath exponent is to 1, the higher the multipath exponent of the multipath pseudopoint. The larger the value, the higher the weight. It exhibits an exponential decay approaching 0.

[0019] S5. The improved weighted DBSCAN algorithm is used to perform clustering region query and expansion on the dynamic adaptive parameters and point cloud weights to generate clusters; S501, Collection Point The expression for calculating the weighted effective number of points in the neighborhood point set is as follows: ; in, Point , neighborhood points; S502, if satisfied , will point Marked as core points and expanded into clusters, where Point The minimum number of points in the dynamic range; otherwise, the points... Mark it as a transient point.

[0020] S6. Based on the consistency between the mean radar cross section and the velocity variance of the cluster, the cluster is filtered for false alarms, outliers of the cluster are removed and marked as final noise, and the point cloud set is obtained. S601, For each cluster Perform internal consistency checks; S602, Calculate the mean within each cluster. and velocity variance ; S603, if the points in the cluster... satisfy Then the decision point Outliers are marked as final noise and removed. Point of Preset deviation threshold (unit: dBsm). This represents the preset velocity variance threshold (unit: (m / s)). ), The intersection is used to represent two conditions that must be met simultaneously; when a point... When the value differs too much from the average level within the cluster, and the speed consistency of the entire cluster is poor, the point is identified as a false alarm and removed. S604. Traverse all points in the cluster to obtain the final point cloud cluster set; like Figure 2The comparative experimental results show that, in typical scenarios where the vehicle's lateral point cloud is sparse and there is multipath interference, the standard DBSCAN (left image) misclassifies the same vehicle target into two independent clusters (red and pink) due to parameter fixation and circular neighborhood limitations. The method provided in this embodiment (right image) successfully clusters the vehicle point cloud into one class (pink) by dynamically relaxing the long-distance parameters and using the ellipsoidal neighborhood to stretch the search range along the main direction. This effectively verifies the outstanding advantages of this invention in solving the problem of undersegmentation of targets and improving the integrity of clustering.

[0021] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A dynamic adaptive anisotropic clustering method for millimeter-wave radar point clouds, characterized in that, Includes the following steps: S1. Data preprocessing and parameter initialization; Input a frame of millimeter-wave radar point cloud data, including each data point Spatial coordinates in the radar coordinate system radial velocity and radar cross section Calculate the radial distance of each data point. ; Regarding the radar cross section Below the preset threshold The data points are preprocessed to filter false alarms and marked as noise points, and then removed from the set of points to be clustered. S2. Calculate the dynamic adaptive parameters, including the dynamic neighborhood radius. and dynamic minimum number of points ; S3. Construct an ellipsoidal neighborhood based on principal component analysis: Traverse every unvisited data point. With the data points Principal component analysis is performed on the center and its local neighborhood to obtain the main direction of the point cloud distribution and to construct an ellipsoidal anisotropic neighborhood. S301, Local Principal Component Analysis: Based on the dynamic neighborhood radius... Define the initial neighborhood and obtain the point set within the neighborhood. And calculate the covariance matrix. ; in, This represents the geometric center of all points within the neighborhood, i.e., the cluster center. Indicates matrix transpose; For the covariance matrix Perform eigenvalue decomposition to obtain eigenvalues. , and the corresponding feature vectors , And satisfy ,in Used to indicate the main direction of point cloud distribution; S302. Define the length of the semi-axis of the ellipsoid, align the major semi-axis with the principal direction, and construct an anisotropic neighborhood of the ellipsoid. S303, Calculate candidate neighborhood points With the center point The offset vectors are respectively along the feature vectors Projection components of direction , Projection components of direction If satisfied Then determine the candidate neighborhood points. Belongs to the center point The ellipsoidal neighborhood; S4, Utilizing radial velocity and radar cross section The multipath evaluation function is used to calculate the multipath index and the point cloud weights. S5. The improved weighted DBSCAN algorithm is used to perform clustering region query and expansion on the dynamic adaptive parameters and point cloud weights to generate clusters; S6. Based on the consistency between the mean radar cross section and the velocity variance of the cluster, the cluster is filtered for false alarms, outliers of the cluster are removed and marked as final noise, and the point cloud set is obtained.

2. The method for dynamic adaptive anisotropic clustering of millimeter-wave radar point clouds according to claim 1, characterized in that, radial distance The expression is: A dynamic adaptive anisotropic clustering method for millimeter-wave radar point clouds according to claim 1, characterized in that the dynamic neighborhood radius is calculated. The expression is: in, Indicates the preset basic neighborhood radius. The expansion factor represents the rate of increase of the control radius with distance. This indicates the preset maximum detection range. Represents the normalized distance factor. ; Calculate the dynamic minimum number of points The expression is: in, This represents the preset minimum number of neighborhood points. This represents the adjustment factor that controls the rate at which the minimum number of points increases with distance.

3. The method for dynamic adaptive anisotropic clustering of millimeter-wave radar point clouds according to claim 1, characterized in that, The expressions for the semi-major and semi-minor axes of the ellipsoid are: in, Represents the semi-major axis of the ellipsoid. Represents the minor semi-axis of the ellipsoid. This represents the preset regularization parameter, used to control the sensitivity of the ellipsoid to anisotropy. This indicates the energy distribution ratio along the principal component direction.

4. The method for dynamic adaptive anisotropic clustering of millimeter-wave radar point clouds according to claim 1, characterized in that, The multipath evaluation function calculation point Multipath index The expression is: in, Point radial velocity, Point The mean radial velocity of all points within the local neighborhood. Point The standard deviation of radial velocity at all points within a local neighborhood. Point radar cross section, This indicates a preset reference scattering cross section based on the target type. , All represent weighting coefficients; When the multipath index Greater than the preset threshold Judgment point Due to multipath interference, attenuation weights are introduced. The expression is: in, This represents the attenuation factor.

5. The method for dynamic adaptive anisotropic clustering of millimeter-wave radar point clouds according to claim 1, characterized in that, The improved weighted DBSCAN algorithm performs clustering region query and expansion on the dynamic adaptive parameters and point cloud weights, specifically including the following steps: S501, Collection Point The number of weighted valid points in the neighborhood point set is calculated using the following expression: in, Point , neighborhood points; S502, if satisfied , will point Marked as core points and expanded into clusters, where Point The minimum number of points in the dynamic range; otherwise, the points... Mark it as a transient point.

6. The method for dynamic adaptive anisotropic clustering of millimeter-wave radar point clouds according to claim 1, characterized in that, Step S6 specifically includes the following steps: S601, For each cluster Perform internal consistency checks; S602, Calculate the mean within each cluster. and velocity variance ; S603, if the points in the cluster... satisfy Then the decision point Outliers are marked as final noise and removed. Point of Preset deviation threshold This represents the preset speed variance threshold. Indicates intersection; S604. Traverse all points in the cluster to obtain the final point cloud cluster set.

Images