Radar point cloud dynamic and static target differentiation detection method, device and system

Abstract

The application relates to the technical field of radar point cloud target detection, and specifically discloses a radar point cloud dynamic and static target differential detection method, device and system, which comprises the following steps: acquiring radar point cloud data; performing format self-adaptive conversion processing on the radar point cloud data; performing ground point cloud filtering processing on the radar point cloud data after format conversion to obtain non-ground effective target point cloud data; performing dynamic and static target segmentation on the non-ground effective target point cloud data according to a multi-dimensional dynamic and static segmentation algorithm including speed error self-adaptive compensation to obtain dynamic point cloud data and static point cloud data; performing point cloud distribution feature mining on the dynamic point cloud data and the static point cloud data respectively, and performing self-adaptive differential clustering processing on the mined dynamic point cloud distribution features and static point cloud distribution features to obtain dynamic target clustering results and static target clustering results. The radar point cloud dynamic and static target differential detection method provided by the application improves target detection precision.

Description

Technical Field

[0001] This invention relates to the field of radar point cloud target detection technology, and in particular to a method, device, and ROS-based radar point cloud target detection system for detecting dynamic and static targets in radar point clouds. Background Technology

[0002] As a core sensor for environmental perception, radar (millimeter-wave radar, lidar) has become a key perception component in fields such as autonomous driving, intelligent security, and mobile robot navigation due to its inherent characteristics of resisting interference from harsh environments such as light, rain, and fog, and being able to directly output the radial velocity of the target.

[0003] Radar point cloud data contains four-dimensional information: the target's three-dimensional spatial coordinates (x, y, z) and radial velocity. It serves as the fundamental data source for target identification, tracking, and environmental modeling. However, raw radar point clouds have inherent limitations: a high proportion of ground-based point clouds, leading to strong interference; overlapping distribution of point clouds for both dynamic and static targets; discrete and fragmented spatial distribution of point clouds for dynamic targets (moving vehicles, pedestrians); and dense spatial distribution of point clouds for static targets (fixed buildings, guardrails, roadblocks), prone to excessive aggregation. Directly relying on raw point clouds for target detection makes accurate identification of effective targets difficult; targeted point cloud preprocessing, segmentation, and clustering analysis are essential for target extraction.

[0004] Existing radar point cloud target detection solutions, such as traditional hard-coded solutions, are based on fixed code logic. The core process involves subscribing to radar point cloud data, ground filtering with fixed thresholds, DBSCAN clustering with fixed parameters, and result output. All core parameters (ground filtering threshold, dynamic / static velocity distinction threshold, clustering eps, min_samples, etc.) are hard-coded into the source code, making it only compatible with a single radar model and a fixed application scenario. This solution suffers from extremely poor parameter configuration flexibility, high scenario adaptation costs, lack of targeted dynamic / static detection algorithms, inherent accuracy defects, high system coupling, poor scalability and maintainability, lack of visualization and debugging capabilities, failure to meet expected functional safety requirements, and weak robustness, making it prone to operational interruptions. For example, the radar point cloud DBSCAN detection scheme based on a single clustering parameter adds dynamic and static point cloud segmentation based on a fixed velocity threshold to the existing technology. However, it uses the same set of DBSCAN clustering parameters for both the segmented dynamic and static point clouds, performing full-dimensional clustering based on the x, y, and z three-dimensional coordinates, and finally outputs the detection results. At the algorithm implementation level, this scheme only uses a fixed velocity threshold for dynamic and static segmentation, without considering the inherent error of radar velocity measurement, and has no discrimination logic for point clouds in velocity ambiguity intervals; the clustering process does not explore the distribution characteristics of dynamic and static point clouds, but only sets fixed parameters based on empirical values, without any adaptive adjustment capability. This scheme suffers from the following problems: the dynamic and static segmentation algorithm is coarse and has inherent segmentation errors; the clustering strategy is unreasonable, resulting in inherent defects in detection accuracy; the clustering dimensions are redundant, introducing additional interference; there is no differentiated result display, resulting in low target recognition; the algorithm lacks adaptive capability and has poor scene adaptability.

[0005] In summary, current mainstream radar point cloud target detection technologies all suffer from technical defects such as poor parameter adaptability, coarse algorithm logic, low clustering accuracy, high debugging difficulty, and weak system robustness. In particular, at the core dynamic and static target detection algorithm level, existing solutions have only made simple procedural designs without deeply exploring the essential characteristics of point clouds of dynamic and static targets. They cannot meet the engineering application requirements of high precision, high adaptability, and high reliability in industrial scenarios, and are also difficult to meet the relevant requirements of ISO21448 for the traceability and debuggability of perception algorithms for expected functional safety.

[0006] Therefore, how to deeply explore the essential features of point clouds of dynamic and static targets to overcome the technical defects of current target point cloud detection, such as segmentation errors caused by coarse algorithm logic and low detection accuracy caused by low clustering accuracy, has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] This invention provides a method, device, and ROS-based radar point cloud target detection system for detecting dynamic and static targets in radar point clouds. It aims to solve at least one of the technical problems in target point cloud detection in related technologies, such as segmentation errors caused by coarse algorithm logic and low detection accuracy caused by low clustering accuracy.

[0008] As a first aspect of the present invention, a method for differential detection of dynamic and static targets in radar point clouds is provided, comprising:

[0009] Acquire radar point cloud data;

[0010] The radar point cloud data is subjected to adaptive format conversion processing to obtain radar point cloud data after format conversion.

[0011] The radar point cloud data after format conversion is filtered by ground point cloud to obtain non-ground effective target point cloud data.

[0012] The non-ground effective target point cloud data is segmented into dynamic and static targets using a multi-dimensional dynamic and static segmentation algorithm, including adaptive compensation for velocity error, to obtain dynamic point cloud data and static point cloud data.

[0013] Point cloud distribution features are mined from both dynamic and static point cloud data. Adaptive differential clustering is then applied to the mined dynamic and static point cloud distribution features to obtain dynamic and static target clustering results.

[0014] Furthermore, the non-ground effective target point cloud data is segmented into dynamic and static targets using a multi-dimensional dynamic and static segmentation algorithm, including adaptive compensation for velocity errors, to obtain dynamic point cloud data and static point cloud data, including:

[0015] The reference speed threshold is dynamically adjusted according to the preset speed error compensation coefficient to obtain the dynamic speed threshold and the static speed threshold.

[0016] The non-ground effective target point cloud data is initially segmented based on the dynamic velocity threshold and the static velocity threshold to obtain preliminary segmentation results. The preliminary segmentation results include candidate dynamic point clouds, candidate static point clouds, and fuzzy region point clouds. The candidate dynamic point clouds are the set of point clouds in the non-ground effective target point cloud data whose absolute point cloud velocity is greater than the dynamic velocity threshold. The candidate static point clouds are the set of point clouds in the non-ground effective target point cloud data whose absolute point cloud velocity is less than the static velocity threshold. The fuzzy region point clouds are the set of point clouds in the non-ground effective target point cloud data whose absolute point cloud velocity is greater than or equal to the static velocity threshold and less than or equal to the dynamic velocity threshold.

[0017] A secondary discrimination is performed on the fuzzy region point cloud in the preliminary segmentation result to obtain the secondary discrimination result of the fuzzy region point cloud.

[0018] The preliminary segmentation results and the secondary discrimination results of the fuzzy region point cloud are fused to obtain a dynamic point cloud set and a static point cloud set.

[0019] Furthermore, a secondary discrimination is performed on the fuzzy region point cloud in the preliminary segmentation result to obtain the secondary discrimination result of the fuzzy region point cloud, including:

[0020] For each point cloud in the fuzzy region point cloud, calculate the number of point clouds in the neighborhood centered on that point cloud and with a preset radius.

[0021] If the number of point clouds in the neighborhood is less than the preset neighborhood density threshold, then the point cloud in the fuzzy region point cloud is determined as a candidate dynamic point cloud.

[0022] If the number of point clouds in the neighborhood is not less than the preset neighborhood density threshold, then determine whether the coordinate offset of the point cloud in the blurred region point cloud within a consecutive preset length of time frame is greater than the preset offset threshold.

[0023] If the offset exceeds the preset threshold, then the point cloud in the blurred region point cloud is determined as a candidate dynamic point cloud;

[0024] If the offset is not greater than the preset offset threshold, then the point cloud in the blurred region point cloud is determined as a candidate dynamic point cloud or a candidate static point cloud.

[0025] Furthermore, point cloud distribution feature mining is performed on both dynamic and static point cloud data. Adaptive differential clustering is then applied to the mined dynamic and static point cloud distribution features to obtain dynamic and static target clustering results, including:

[0026] Point cloud distribution features are quantized for dynamic point cloud data and static point cloud data respectively to obtain dynamic point cloud quantization features and static point cloud quantization features.

[0027] The dynamic clustering parameters are iteratively calibrated based on the dynamic point cloud quantization features to obtain calibrated dynamic clustering parameters, and the static clustering parameters are iteratively calibrated based on the static point cloud quantization features to obtain calibrated static clustering parameters.

[0028] The dynamic initial effective clustering results are obtained by density clustering based on the x and y plane coordinates of the point cloud according to the calibrated dynamic clustering parameters, and the static initial effective clustering results are obtained by density clustering based on the x and y plane coordinates of the point cloud according to the calibrated static clustering parameters.

[0029] The dynamic initial valid clustering results and the static initial valid clustering results are filtered and the target attributes are verified twice to obtain the dynamic target clustering results and the static target clustering results.

[0030] Furthermore, point cloud distribution features are quantized for both dynamic and static point cloud data to obtain dynamic point cloud quantization features and static point cloud quantization features, including:

[0031] The point cloud x and y planes in the dynamic point cloud data are divided according to the preset basic sampling interval to obtain multiple sampling grids;

[0032] Calculate the point cloud density within each sampling grid;

[0033] The dynamic point cloud discreteness feature value is determined based on the standard deviation of the point cloud density in all sampling grids, and the dynamic point cloud discreteness feature value is determined as the dynamic point cloud quantization feature.

[0034] For each point in the static point cloud data, calculate its average distance to K nearest neighbors, where K = min_s_s0, min_s_s0 represents the basic min_samples value of static clustering, and min_samples represents the minimum number of data points required for an effective cluster;

[0035] The clustering feature value is obtained by taking the mean of the average distances corresponding to all points in the static point cloud data and then calculating the reciprocal of the mean.

[0036] The aggregation feature value is determined as the static point cloud quantization feature.

[0037] Further, the process of iteratively calibrating the dynamic clustering parameters based on the dynamic point cloud quantization features to obtain calibrated dynamic clustering parameters, and iteratively calibrating the static clustering parameters based on the static point cloud quantization features to obtain calibrated static clustering parameters, includes:

[0038] Dynamic clustering parameters are determined based on the dynamic point cloud quantization features and the preset dynamic point cloud standard discreteness feature values, and the dynamic clustering parameters are calibrated. The expression for dynamic clustering parameter calibration is as follows:

[0039] eps_d = eps_d0 × (1 + 0.5×(ρ_d0 / ρ_d - 1)),

[0040] min_samples_d = max(round(min_s_d0 × (ρ_d0 / ρ_d)), 1),

[0041] Where eps_d represents the dynamic clustering parameters, eps_d0 represents the basic eps value of dynamic clustering, ρ_d0 represents the preset standard discreteness feature value of dynamic point cloud, ρ_d represents the discreteness feature value of dynamic point cloud, min_samples_d represents the minimum number of data points required for a valid cluster in dynamic clustering, min_s_d0 represents the basic min_samples value of dynamic clustering, and eps represents the neighborhood search radius of the clustering algorithm;

[0042] Static clustering parameters are determined based on the static point cloud quantization features and preset static point cloud standard clustering feature values, and the static clustering parameters are calibrated. The expression for static clustering parameter calibration is as follows:

[0043] eps_s = eps_s0 × (1 - 0.3×(α_s / α_s0 - 1)),

[0044] min_samples_s = max(round(min_s_s0 × (α_s / α_s0)), 3),

[0045] Where eps_s represents the static clustering parameters, eps_s0 represents the basic eps value of static clustering, α_s0 represents the preset standard clustering feature value of static point cloud, α_s represents the static point cloud quantization feature, min_samples_s represents the minimum number of data points required for a valid cluster in static clustering, and min_s_s0 represents the basic min_samples value of static clustering.

[0046] The calibration process for both dynamic and static clustering parameters is iterated a preset number of times, and the clustering effect evaluation index is calculated after each iteration.

[0047] If the difference between the clustering performance evaluation index of two consecutive iterations is less than the preset convergence threshold, the iteration is determined to be converged and the iteration is stopped. The dynamic clustering parameters and static clustering parameters are output.

[0048] If the iteration fails to converge after a preset number of iterations, the calibration results corresponding to the dynamic clustering basic parameters and the static clustering basic parameters will be output respectively.

[0049] Further, the radar point cloud data undergoes adaptive format conversion processing to obtain format-converted radar point cloud data, including:

[0050] The radar point cloud data is automatically adapted to the point cloud field structure, and the four-dimensional valid fields after the field structure adaptation are read.

[0051] Filter outliers in the velocity field among the valid fields in the four-dimensional array;

[0052] Convert the four-dimensional valid fields after outlier filtering into a preset unified data format.

[0053] Furthermore, the converted radar point cloud data undergoes ground point cloud filtering to obtain non-ground effective target point cloud data, including:

[0054] The radar point cloud data after format conversion is subjected to preliminary ground point cloud filtering processing according to the preset ground point cloud filtering threshold range to obtain non-ground candidate point clouds.

[0055] Normal vector estimation is performed on the non-ground candidate point cloud to perform secondary filtering of the ground point cloud based on the angle between the normal vector of each point cloud in the non-ground candidate point cloud and the ground normal vector, so as to obtain the non-ground effective target point cloud data.

[0056] As another aspect of the present invention, a radar point cloud dynamic and static target differentiation detection device is provided for implementing the radar point cloud dynamic and static target differentiation detection method described above, wherein the device includes:

[0057] The acquisition module is used to acquire radar point cloud data;

[0058] The format conversion module is used to perform adaptive format conversion processing on the radar point cloud data to obtain the format-converted radar point cloud data.

[0059] The point cloud filtering module is used to perform ground point cloud filtering on the radar point cloud data after format conversion to obtain non-ground effective target point cloud data.

[0060] The segmentation module is used to perform dynamic and static target segmentation on the non-ground effective target point cloud data according to a multi-dimensional dynamic and static segmentation algorithm including adaptive compensation for velocity error, so as to obtain dynamic point cloud data and static point cloud data.

[0061] The clustering module is used to mine point cloud distribution features of dynamic point cloud data and static point cloud data respectively, and to perform adaptive differential clustering processing on the mined dynamic point cloud distribution features and static point cloud distribution features to obtain dynamic target clustering results and static target clustering results.

[0062] As another aspect of the present invention, a radar point cloud target detection system based on ROS is provided, comprising: a radar sensing device and the radar point cloud dynamic and static target differentiation detection device described above, wherein the radar sensing device is communicatively connected to the radar point cloud dynamic and static target differentiation detection device.

[0063] The radar sensing device is used to collect radar point cloud data;

[0064] The radar point cloud dynamic and static target differentiation detection device is used to perform dynamic and static differentiation processing on the radar point cloud data to obtain dynamic target clustering results and static target clustering results.

[0065] The radar point cloud dynamic and static target differentiation detection method provided by this invention acquires radar point cloud data, performs adaptive format conversion on the radar point cloud data, filters the converted radar point cloud data as ground point clouds, segments non-ground effective target point cloud data into dynamic and static targets using a multi-dimensional dynamic and static segmentation algorithm, mines point cloud distribution features from both dynamic and static point cloud data, and performs adaptive differential clustering processing on the mined dynamic and static point cloud distribution features to obtain dynamic target clustering results and static target clustering results. This radar point cloud dynamic and static target differentiation detection method effectively compensates for the inherent error of radar velocity measurement through a multi-dimensional accurate segmentation algorithm, solving the defect of misjudgment of ambiguous regions caused by traditional fixed threshold segmentation. Furthermore, through quantitative analysis of dynamic point cloud dispersion and static point cloud clustering, online adaptive adjustment of clustering parameters is achieved, avoiding the inherent defect of traditional fixed parameters failing to match point cloud distribution differences, effectively improving target detection accuracy. Therefore, the radar point cloud dynamic and static target differentiation detection method provided by the present invention can effectively solve the problems of segmentation error caused by coarse logic in existing algorithms and low detection accuracy caused by low clustering accuracy, thereby improving the target detection accuracy. Attached Figure Description

[0066] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the following detailed description to explain the invention, but do not constitute a limitation thereof.

[0067] Figure 1 The flowchart is for the radar point cloud dynamic and static target differentiation detection method provided by the present invention.

[0068] Figure 2 This is a flowchart for adaptive format conversion of radar point cloud data provided by the present invention.

[0069] Figure 3 This is a flowchart of ground point cloud filtering processing provided by the present invention.

[0070] Figure 4 The flowchart for dynamic and static target segmentation provided by the present invention.

[0071] Figure 5 The flowchart for the adaptive differential clustering process provided by this invention is shown.

[0072] Figure 6 The structural block diagram of the radar point cloud dynamic and static target differentiation detection device provided by the present invention is shown.

[0073] Figure 7 The structural block diagram of the ROS-based radar point cloud target detection system provided by this invention. Detailed Implementation

[0074] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0075] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0076] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of the invention described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0077] This embodiment provides a method for detecting the differences between static and dynamic targets in radar point clouds. Figure 1 This is a flowchart of a radar point cloud dynamic and static target differentiation detection method provided by an embodiment of the present invention, such as... Figure 1 As shown, it includes:

[0078] S100, Acquire radar point cloud data;

[0079] In this embodiment of the invention, the ROS (Robot Operating System) subscription and publish mechanism can be used to subscribe to ROS PointCloud2 format radar point cloud data in the input topic, and the subsequent processing flow can be triggered immediately after receiving the radar point cloud data.

[0080] S200: Perform adaptive format conversion on the radar point cloud data to obtain radar point cloud data after format conversion;

[0081] In this embodiment of the invention, the radar point cloud data is format-converted and automatically adapted to the point cloud field structure to obtain the format-converted radar point cloud data.

[0082] S300: Perform ground point cloud filtering on the radar point cloud data after format conversion to obtain non-ground effective target point cloud data;

[0083] In this embodiment of the invention, the radar point cloud data after format conversion is subjected to preliminary ground filtering and secondary ground filtering to obtain non-ground effective target point cloud data.

[0084] S400. Perform dynamic and static target segmentation on the non-ground effective target point cloud data according to a multi-dimensional dynamic and static segmentation algorithm including adaptive compensation for velocity error, to obtain dynamic point cloud data and static point cloud data.

[0085] In this embodiment of the invention, based on velocity information and combined with spatial distribution characteristics and multi-frame temporal consistency, a multi-dimensional dynamic and static segmentation algorithm is designed to achieve accurate segmentation of dynamic and static targets in non-ground effective target point cloud data.

[0086] S500 performs point cloud distribution feature mining on dynamic point cloud data and static point cloud data respectively, and performs adaptive differential clustering processing on the mined dynamic point cloud distribution features and static point cloud distribution features to obtain dynamic target clustering results and static target clustering results.

[0087] In this embodiment of the invention, taking into account the differences in spatial distribution characteristics between dynamic point cloud data and static point cloud data, a combined strategy of point cloud distribution feature quantization, adaptive iterative calibration of clustering parameters, differential clustering in the x and y planes, and optimization of clustering results is adopted to achieve high-precision clustering.

[0088] Therefore, the radar point cloud dynamic and static target differentiation detection method provided by this invention acquires radar point cloud data, performs adaptive format conversion on the radar point cloud data, performs ground point cloud filtering on the converted radar point cloud data, performs dynamic and static target segmentation on non-ground effective target point cloud data according to a multi-dimensional dynamic and static segmentation algorithm, and mines point cloud distribution features for both dynamic and static point cloud data. Finally, it performs adaptive differential clustering processing on the mined dynamic and static point cloud distribution features to obtain dynamic target clustering results and static target clustering results. This radar point cloud dynamic and static target differentiation detection method can effectively compensate for the inherent error of radar velocity measurement through a multi-dimensional accurate segmentation algorithm, solving the defect of misjudgment of ambiguous areas caused by traditional fixed threshold segmentation. In addition, through quantitative analysis of dynamic point cloud dispersion and static point cloud clustering, it achieves online adaptive adjustment of clustering parameters, avoiding the essential defect of traditional fixed parameters being unable to match point cloud distribution differences, and effectively improving target detection accuracy. Therefore, the radar point cloud dynamic and static target differentiation detection method provided by the present invention can effectively solve the problems of segmentation error caused by coarse logic in existing algorithms and low detection accuracy caused by low clustering accuracy, thereby improving the target detection accuracy.

[0089] In this embodiment of the invention, the radar point cloud data undergoes adaptive format conversion processing to obtain format-converted radar point cloud data, such as... Figure 2 As shown, it includes:

[0090] S210. Automatically adapt the radar point cloud data to the point cloud field structure and read the four-dimensional valid fields after the field structure adaptation.

[0091] Specifically, the received point cloud data is formatted and automatically adapted to the point cloud field structure, and the four valid fields of x, y, z, and velocity in the point cloud are read.

[0092] S220. Filter outliers in the velocity field among the valid fields of the four dimensions.

[0093] Specifically, invalid point clouds containing NaN, Inf values, and velocity anomalies are filtered out using the 3σ principle.

[0094] In this embodiment of the invention, the outlier filtering of the velocity field adopts the 3σ principle, and the specific logic is as follows:

[0095] (1) Calculate the mean μ and standard deviation σ of the total effective point cloud velocity values;

[0096] (2) Remove abnormal point clouds with velocity values ​​outside the range of [μ-3σ, μ+3σ] to avoid interference from abnormal velocity values ​​on subsequent dynamic and static segmentation;

[0097] (3) For the filtered point cloud, construct an N×4 NumPy array according to the four dimensions of x, y, z, and velocity, where N is the number of valid point clouds.

[0098] S230. Convert the four-dimensional valid fields after outlier filtering into a preset unified data format.

[0099] Specifically, the point cloud data is converted into NumPy array format to provide a unified data format for subsequent numerical calculations; if the point cloud is missing the velocity field, the velocity information is assigned according to the configured default value.

[0100] In this embodiment of the invention, the radar point cloud data after format conversion is subjected to ground point cloud filtering processing to obtain non-ground effective target point cloud data, such as... Figure 3 As shown, it includes:

[0101] S310. Perform preliminary ground point cloud filtering on the radar point cloud data after format conversion according to the preset ground point cloud filtering threshold range to obtain non-ground candidate point clouds.

[0102] Specifically, ground point clouds with z-coordinates within the threshold range of [z_min, z_max] are removed by logical judgment, while non-ground candidate point clouds are retained.

[0103] S320. Perform normal vector estimation on the non-ground candidate point cloud to perform secondary filtering of the ground point cloud based on the angle between the normal vector of each point cloud in the non-ground candidate point cloud and the ground normal vector, and obtain non-ground effective target point cloud data.

[0104] Specifically, normal vector estimation is performed on the candidate point cloud, the angle between the normal vector of each point cloud and the ground normal vector (positive z-axis direction) is calculated, and the remaining ground point clouds with angles less than the configured threshold (default 15°, which can be adjusted externally) are removed, and finally the non-ground effective target point cloud is obtained.

[0105] In this embodiment of the invention, in order to provide data reference for subsequent parameter optimization, the original radar point cloud count, the initial filtered point cloud count, and the final effective point cloud count can be output to the ROS log.

[0106] In this embodiment of the invention, the non-ground effective target point cloud data is segmented into dynamic and static targets according to a multi-dimensional dynamic and static segmentation algorithm including adaptive compensation for velocity errors, to obtain dynamic point cloud data and static point cloud data, such as... Figure 4 As shown, it includes:

[0107] S410. Dynamically adjust the reference speed threshold according to the preset speed error compensation coefficient to obtain the dynamic speed threshold and the static speed threshold.

[0108] Specifically, based on the speed measurement error characteristics of the radar sensor, a speed error compensation coefficient k_v is introduced to dynamically adjust the reference speed threshold V_base, resulting in a dynamic speed threshold V_d and a static speed threshold V_s: dynamic speed threshold V_d = V_base × (1 + k_v), static speed threshold V_s = V_base × (1 - k_v). The speed error compensation coefficient k_v can be configured externally through a launch file according to the measurement accuracy of the radar model. The recommended value range is 0.1~0.3, which is used to compensate for the segmentation deviation caused by the radar speed measurement error and avoid missegmentation caused by a fixed threshold.

[0109] It should be understood that the speed error compensation coefficient is an adaptive coefficient used to correct the inherent error of radar speed measurement. It is obtained based on the measurement accuracy calibration of the radar model and realizes the dynamic adjustment of the dynamic and static segmentation threshold, which conforms to the general specifications for vehicle-mounted radar measurement error correction.

[0110] S420. Perform preliminary segmentation on the non-ground effective target point cloud data according to the dynamic velocity threshold and the static velocity threshold to obtain preliminary segmentation results. The preliminary segmentation results include candidate dynamic point clouds, candidate static point clouds, and fuzzy region point clouds. The candidate dynamic point cloud is the set of point clouds in the non-ground effective target point cloud data whose absolute point cloud velocity is greater than the dynamic velocity threshold. The candidate static point cloud is the set of point clouds in the non-ground effective target point cloud data whose absolute point cloud velocity is less than the static velocity threshold. The fuzzy region point cloud is the set of point clouds in the non-ground effective target point cloud data whose absolute point cloud velocity is greater than or equal to the static velocity threshold and less than or equal to the dynamic velocity threshold.

[0111] Specifically, based on the adjusted velocity threshold, the non-ground effective point cloud is initially segmented into three types of point cloud sets:

[0112] 1) Candidate dynamic point cloud P_d: The set of point clouds with absolute point cloud velocity > V_d, corresponding to high-confidence dynamic targets;

[0113] 2) Candidate static point cloud P_s: The set of point clouds with absolute point cloud velocity values ​​< V_s, corresponding to high-confidence static targets;

[0114] 3) Blurred region point cloud P_m: The set of point clouds whose absolute velocity values ​​are within the interval [V_s, V_d], corresponding to the low-confidence point cloud in the fuzzy region of velocity measurement.

[0115] S430. Perform secondary discrimination on the fuzzy region point cloud in the preliminary segmentation result to obtain the secondary discrimination result of the fuzzy region point cloud;

[0116] For the point cloud P_m in the fuzzy region, a two-dimensional feature of "neighborhood density feature + multi-frame temporal consistency" is introduced for secondary discrimination.

[0117] Specifically, a secondary discrimination is performed on the fuzzy region point cloud in the preliminary segmentation result to obtain the secondary discrimination result of the fuzzy region point cloud, including:

[0118] (1) For each point cloud in the fuzzy region point cloud, calculate the number of point clouds in the neighborhood centered on the point cloud and with a preset radius;

[0119] In this embodiment of the invention, for each point cloud in P_m, the number of point clouds ρ (neighborhood density) in the neighborhood centered on that point and with a configuration radius (default r=0.5m, which can be adjusted externally) is calculated.

[0120] (2) If the number of point clouds in the neighborhood is less than the preset neighborhood density threshold, then the point cloud in the fuzzy region point cloud is determined as a candidate dynamic point cloud;

[0121] In this embodiment of the invention, if ρ < ρ_th (ρ_th represents the density threshold, which can be configured externally), it is determined to be a candidate dynamic point cloud and is included in the candidate dynamic point cloud P_d.

[0122] (3) If the number of point clouds in the neighborhood is not less than the preset neighborhood density threshold, then determine whether the coordinate offset of the point cloud in the blurred area point cloud within a consecutive preset length of time frame is greater than the preset offset threshold.

[0123] In this embodiment of the invention, if ρ ≥ ρ_th, then the multi-frame temporal consistency determination stage is entered. Specifically, in the multi-frame temporal consistency determination, a temporal sliding window of length T is constructed, and the coordinate offset Δd of the point cloud within consecutive T frames is calculated.

[0124] (4) If the offset is greater than the preset offset threshold, then the point cloud in the blurred region point cloud is determined as a candidate dynamic point cloud;

[0125] In this embodiment of the invention, if Δd > d_th (d_th represents the offset threshold, which can be configured externally), it is determined to be a dynamic point cloud and included in the candidate dynamic point cloud P_d.

[0126] (5) If it is not greater than the preset offset threshold, then the point cloud in the fuzzy area point cloud is determined as a candidate dynamic point cloud or a candidate static point cloud.

[0127] In this embodiment of the invention, if Δd≤d_th, it is determined to be a static point cloud and is included in the candidate static point cloud P_s.

[0128] S440. The preliminary segmentation results and the secondary discrimination results of the fuzzy region point cloud are fused to obtain a dynamic point cloud set and a static point cloud set.

[0129] In this embodiment of the invention, the preliminary segmentation result is fused with the secondary discrimination result of the fuzzy region to obtain the final dynamic point cloud set D and static point cloud set S. The dynamic point cloud set D is a candidate dynamic point cloud P_d combined with the point cloud determined to be dynamic in the fuzzy region; the static point cloud set S is a candidate static point cloud P_s combined with the point cloud determined to be static in the fuzzy region.

[0130] In this embodiment of the invention, after segmentation is completed, confidence verification is performed. Specifically, the segmentation confidence C is calculated using the following formula:

[0131] C = (Percentage of velocity-consistent point cloud in D + Percentage of density-consistent point cloud in S) / 2

[0132] If C < the configured threshold (default 0.85, can be adjusted externally), a warning log will be output, prompting the user to optimize the speed threshold or density threshold parameters;

[0133] Specifically, the number of dynamic point clouds, the number of static point clouds, and the segmentation confidence can be output to the ROS log to achieve traceability of the target detection process.

[0134] In this embodiment of the invention, point cloud distribution feature mining is performed on dynamic point cloud data and static point cloud data respectively, and adaptive differential clustering processing is applied to the mined dynamic point cloud distribution features and static point cloud distribution features to obtain dynamic target clustering results and static target clustering results, such as... Figure 5 As shown, it includes:

[0135] S510. Perform point cloud distribution feature quantization on dynamic point cloud data and static point cloud data respectively to obtain dynamic point cloud quantization features and static point cloud quantization features.

[0136] In this embodiment of the invention, the distribution characteristics of dynamic point clouds and static point clouds are calculated respectively to explore their inherent distribution characteristics and provide a basis for clustering parameter calibration.

[0137] Specifically, point cloud distribution features are quantized for both dynamic and static point cloud data to obtain dynamic point cloud quantization features and static point cloud quantization features, including:

[0138] (1) Divide the x and y planes of the point cloud in the dynamic point cloud data according to the preset basic sampling interval to obtain multiple sampling grids;

[0139] Specifically, the x and y planes of the dynamic point cloud are divided into multiple sampling grids based on the point cloud density sampling interval d_sample.

[0140] (2) Calculate the point cloud density within each sampling grid;

[0141] (3) Determine the dynamic point cloud discreteness feature value based on the standard deviation of the point cloud density in all sampling grids, and determine the dynamic point cloud discreteness feature value as the dynamic point cloud quantization feature;

[0142] In this embodiment of the invention, the standard deviation of all grid densities is taken as the dynamic point cloud dispersion characteristic value ρ_d. The larger the dynamic point cloud dispersion characteristic value ρ_d is, the more discrete the dynamic point cloud distribution is.

[0143] It should be noted that the point cloud discreteness feature value can be understood as a numerical index that quantifies the degree of dispersion of the dynamic point cloud spatial distribution. It is calculated by the standard deviation of the neighborhood point cloud density through multi-scale grid sampling and is used for adaptive calibration of dynamic clustering parameters.

[0144] (4) For each point cloud in the static point cloud data, calculate the average distance to its K nearest neighbors, where K = min_s_s0, min_s_s0 represents the basic min_samples value of static clustering, and min_samples represents the minimum number of data points required for an effective cluster;

[0145] In this embodiment of the invention, the K-nearest neighbor algorithm is used to calculate the average distance from each point in the static point cloud to its K=min_s_s0 nearest neighbor points.

[0146] (5) Take the mean of the average distances corresponding to all points in the static point cloud data and then calculate the reciprocal to obtain the clustering feature value;

[0147] In this embodiment of the invention, the reciprocal of the mean of the average distance of all points is taken to obtain the clustering characteristic value α_s. The larger α_s is, the more concentrated the static point cloud distribution is.

[0148] It should be noted that the clustering feature value can be understood as a numerical index that quantifies the degree of concentration of the spatial distribution of static point clouds. It is obtained by calculating the inverse of the average distance of the point clouds within the cluster using the K-nearest neighbor algorithm and is used for adaptive calibration of static clustering parameters.

[0149] (6) The aggregation feature value is determined as the static point cloud quantization feature.

[0150] S520. Iteratively calibrate the dynamic clustering parameters according to the dynamic point cloud quantization features to obtain calibrated dynamic clustering parameters, and iteratively calibrate the static clustering parameters according to the static point cloud quantization features to obtain calibrated static clustering parameters.

[0151] In this embodiment of the invention, based on the quantized distribution characteristics, the basic clustering parameters are iteratively calibrated online to ensure that the parameters accurately match the point cloud distribution characteristics.

[0152] Specifically, the process involves iteratively calibrating the dynamic clustering parameters based on the dynamic point cloud quantization features to obtain calibrated dynamic clustering parameters, and iteratively calibrating the static clustering parameters based on the static point cloud quantization features to obtain calibrated static clustering parameters, including:

[0153] (1) Determine the dynamic clustering parameters based on the dynamic point cloud quantization features and the preset dynamic point cloud standard discreteness feature value, and calibrate the dynamic clustering parameters, wherein the expression for dynamic clustering parameter calibration is:

[0154] eps_d = eps_d0 × (1 + 0.5×(ρ_d0 / ρ_d - 1)),

[0155] min_samples_d = max(round(min_s_d0 × (ρ_d0 / ρ_d)), 1),

[0156] Where eps_d represents the dynamic clustering parameters, eps_d0 represents the basic eps value of dynamic clustering, ρ_d0 represents the preset standard discreteness feature value of dynamic point cloud, ρ_d represents the discreteness feature value of dynamic point cloud, min_samples_d represents the minimum number of data points required for a valid cluster in dynamic clustering, min_s_d0 represents the basic min_samples value of dynamic clustering, and eps represents the neighborhood search radius of the clustering algorithm;

[0157] It should be noted that the preset dynamic point cloud standard discreteness feature value ρ_d0 can be pre-configured. The smaller the dynamic point cloud discreteness feature value ρ_d (the more discrete the point cloud), the larger the dynamic clustering parameter eps_d and the smaller the min_samples_d, which adapts to the discrete distribution characteristics of dynamic targets and thus fundamentally avoids fragmentation. It should also be noted that eps is a core parameter of the DBSCAN clustering algorithm, referring to the neighborhood search radius, used to determine whether two data points belong to the same neighborhood.

[0158] (2) Determine the static clustering parameters based on the static point cloud quantization features and the preset static point cloud standard clustering feature value, and calibrate the static clustering parameters, wherein the expression for static clustering parameter calibration is:

[0159] eps_s = eps_s0 × (1 - 0.3×(α_s / α_s0 - 1)),

[0160] min_samples_s = max(round(min_s_s0 × (α_s / α_s0)), 3),

[0161] Where eps_s represents the static clustering parameters, eps_s0 represents the basic eps value of static clustering, α_s0 represents the preset standard clustering feature value of static point cloud, α_s represents the static point cloud quantization feature, min_samples_s represents the minimum number of data points required for a valid cluster in static clustering, and min_s_s0 represents the basic min_samples value of static clustering.

[0162] It should be noted that the preset static point cloud standard clustering feature value α_s0 can be configured in advance; the larger the static point cloud quantization feature α_s (the more concentrated the point cloud), the smaller the static clustering parameter eps_s and the larger the min_samples_s, which adapts to the concentrated distribution characteristics of static targets and fundamentally avoids over-clustering.

[0163] (3) The calibration process for both dynamic and static clustering parameters is iterated a preset number of times, and the clustering effect evaluation index is calculated after each iteration;

[0164] In this embodiment of the invention, the calibration process is iterated a preset number of times N_iter, and after each iteration, the clustering effect evaluation index Q (number of effective clusters / total number of point clouds) is calculated.

[0165] (4) If the difference between the clustering effect evaluation index of two adjacent iterations is less than the preset convergence threshold, the iteration is determined to be converged and the iteration is stopped. The dynamic clustering parameters and static clustering parameters are output.

[0166] Specifically, if the difference in Q-value between two consecutive iterations is less than the configured convergence threshold (default 0.05, which can be adjusted externally), the iteration stops and the final clustering parameters are output.

[0167] (5) If the iteration fails to converge after a preset number of iterations, output the calibration results corresponding to the dynamic clustering basic parameters and the static clustering basic parameters respectively.

[0168] Specifically, if the iteration fails to converge after reaching the maximum number of iterations, the calibration result corresponding to the basic parameters is output.

[0169] S530. Based on the calibrated dynamic clustering parameters, density clustering is performed on the x and y plane coordinates of the point cloud to obtain dynamic initial effective clustering results, and based on the calibrated static clustering parameters, density clustering is performed on the x and y plane coordinates of the point cloud to obtain static initial effective clustering results.

[0170] In this embodiment of the invention, dynamic point cloud clustering can specifically use calibrated dynamic clustering parameters (eps_d, min_samples_d) to perform density clustering based solely on the x and y plane coordinates of the point cloud, eliminating irrelevant interference from the z-axis and velocity dimensions; static point cloud clustering can specifically use calibrated static clustering parameters (eps_s, min_samples_s) to perform density clustering based solely on the x and y plane coordinates of the point cloud, eliminating the impact of z-axis dimension interference on clustering accuracy.

[0171] S540. The dynamic initial effective clustering results and the static initial effective clustering results are filtered and the target attributes are verified twice to obtain the dynamic target clustering results and the static target clustering results.

[0172] In this embodiment of the invention, after clustering is completed, noise points marked as -1 by the DBSCAN algorithm are excluded, and only the point cloud set with non-negative integer labels is used as the initial valid cluster. The initial valid clusters are then filtered, and small clusters with a point cloud count < min_cluster_size (which can be configured externally; dynamic clusters default to 3, and static clusters default to 5) are removed to avoid false targets formed by noise points. For the dynamic clustering results, the velocity consistency of the point clouds within the cluster is checked. If the velocity standard deviation is > 2 × V_base, it is determined to be an invalid cluster and is removed. For the static clustering results, the positional stability of the point clouds within the cluster is checked. Combining multi-frame temporal information, if the center point offset is > d_th within T consecutive frames, it is determined to be an invalid cluster and is removed. For each valid cluster, a weighted average method is used to calculate the three-dimensional center point coordinates (x and y coordinates are weighted by point cloud density, and z coordinate is weighted by point cloud height) to reduce the center point calculation bias.

[0173] To achieve differentiated visualization and publishing, after obtaining dynamic and static target clustering results, these results are converted into ROS MarkerArray (a standard message type in the ROS ecosystem used for visualization marking, capable of carrying various geometric shapes, text, arrows, and other marker elements, used for displaying target results in visualization tools such as RViz). This enables differentiated visualization and publishing of the detection results. Specifically:

[0174] (1) Triple differentiation label: The clustering results of dynamic targets are displayed using red sphere markers, and the clustering results of static targets are displayed using blue cube markers; at the same time, text labels are added to each marker to indicate the target type (dynamic / static), cluster confidence, and center point coordinates. Through the triple differentiation of shape, color, and text, the dynamic and static targets can be intuitively distinguished.

[0175] (2) Precise coordinate system matching: The coordinate system of the Marker is consistent with the native frame_id of the radar point cloud to avoid visualization position errors caused by coordinate system deviation;

[0176] (3) Marker attribute optimization: The Marker ID adopts a 0~999 cyclic reuse mechanism to avoid visualization failure caused by ID overflow; the Marker is set to 0.7 transparency to avoid visualization blur caused by target overlap; the Marker is set to 0.5s life cycle to adapt to the high frequency output characteristics of radar and ensure the continuity and clarity of visualization effect.

[0177] (4) Results publishing: Publish the generated MarkerArray data to the pre-configured output topic, supporting real-time viewing by ROS visualization tools such as Rviz (the final long-term support version of ROS 1, adapted to the Ubuntu 20.04 operating system, and the mainstream ROS version used in the industry).

[0178] In this embodiment of the invention, the ground-filtered point cloud obtained in step S300, the intermediate point cloud of dynamic and static segmentation obtained in step S400, and the intermediate point cloud of clustering obtained in step S500 are respectively restored from NumPy array format to the standard ROSPointCloud2 format (the standard message type used in the ROS ecosystem for transmitting 3D point cloud data, which is the core data carrier of 3D perception sensors such as radar and lidar) and published to three independent debugging topics. Operators can add PointCloud2 display components in RViz, subscribe to the above topics, and view the ground filtering effect, dynamic and static segmentation effect, and clustering effect in real time, quickly complete parameter optimization and adjustment, and realize the traceability and debuggability of the entire algorithm process.

[0179] In this embodiment of the invention, exception handling logic is embedded in the entire process described above, specifically including:

[0180] (1) Empty point cloud processing: If an empty point cloud is received in step S200, a warning log is immediately output and all subsequent processing steps are skipped;

[0181] (2) Invalid point cloud processing: If there are no valid point clouds after ground filtering in step S300, immediately output the debugging log and terminate the subsequent segmentation and clustering steps;

[0182] (3) Clustering failure handling: If the number of effective clusters after clustering in step S500 is 0, the convergence threshold of the self-calibration of clustering parameters is automatically reduced and the clustering process is re-executed; if the secondary clustering still fails, a warning log is output and the original point cloud is retained for subsequent analysis.

[0183] (4) Global anomaly capture: Global anomaly capture is performed on the entire processing flow. If anomalies such as data format error, parameter configuration error, or clustering calculation error occur, detailed error logs are immediately output. At the same time, the ROS node is kept running normally to avoid interruption of the entire detection process due to failure of single frame point cloud processing.

[0184] In summary, the radar point cloud dynamic and static target differentiation detection method provided by this invention has the following advantages compared with the prior art:

[0185] (1) Multi-dimensional dynamic and static segmentation algorithm, which solves the inherent segmentation error defects of existing technologies in principle: Specifically, through the multi-dimensional segmentation algorithm of "adaptive adjustment of velocity threshold + secondary discrimination of multiple features in fuzzy regions + multi-frame temporal verification", the inherent error of radar velocity measurement is effectively compensated, and the inherent defect of misjudgment of fuzzy regions caused by traditional fixed threshold segmentation is solved. Based on this technical solution, compared with the existing fixed threshold segmentation scheme, the present invention can effectively reduce the missegmentation rate caused by velocity measurement error, and theoretically the segmentation accuracy can be effectively improved, laying a core foundation for the subsequent improvement of clustering accuracy.

[0186] (2) Adaptive Differential Clustering Algorithm: Addressing Common Industry Pain Points at Their Roots. Specifically, through a four-level combined strategy of "quantification of distribution features, adaptive iterative calibration of parameters, dimensionality reduction clustering, and result optimization," the algorithm achieves online adaptive adjustment of clustering parameters by quantitatively analyzing the dynamic point cloud dispersion and static point cloud clustering, thus avoiding the inherent defect of traditional fixed parameters failing to match the differences in point cloud distribution. At its root, this invention can simultaneously address the common industry pain points of dynamic target fragmentation and static target over-clustering. Compared to existing single-parameter three-dimensional clustering schemes, theoretically, clustering accuracy can be significantly improved, and the deviation in target center point calculation can be greatly reduced.

[0187] (3) All core parameters are externally configurable, and the adaptability of scenarios is significantly improved: All core parameters (including the speed error compensation coefficient, sampling interval, number of iterations, etc. at the algorithm level) are externally configured through the ROS launch file without modifying the internal code of the system. This allows for quick adaptation to the detection needs of different radar models and different application scenarios, solving the industry pain points of hard-coded traditional technical parameters and poor adaptability. At the same time, the parameters are automatically verified and output after reading, which facilitates quick verification of configuration consistency and greatly reduces the difficulty and time cost of on-site debugging and adaptation.

[0188] (4) Modular design throughout the process, significantly enhancing scalability and maintainability: The functions of point cloud reception, format conversion, ground filtering, dynamic and static segmentation, clustering, visualization, and debugging are separated into independent functional modules. Each module is responsible for only a single function, and the modules communicate with each other through standard data formats, resulting in extremely low coupling. Subsequent additions of functions such as target trajectory tracking and classification recognition can be achieved simply by adding new modules without modifying the original logic, which greatly reduces the maintenance and expansion costs of the system.

[0189] (5) The whole process visualization and debugging are integrated, which meets the latest expected functional safety standards: the intuitive visualization of dynamic and static targets is achieved through the triple differentiation of shape, color and text. Operators can view the detection results in real time through RViz. At the same time, three independent debugging point cloud release channels are set up for ground filtering, segmentation and clustering to realize the visualization debugging of the whole process data. Combined with the real-time data statistics of ROS logs, an integrated capability of "detection result verification + process data debugging + whole process traceability" is constructed, which meets the expected functional safety requirements of ISO 21448 (specifically corresponding to expected functional safety SOTIF, which is the safety operation specification for perception systems in complex scenarios in the field of autonomous driving) for the traceability and debuggability of autonomous driving perception systems. The debugging efficiency can be significantly improved compared with traditional solutions.

[0190] (6) Perfected exception handling with robustness reaching industrial application standards: Designed exception handling logic for multiple scenarios such as empty point cloud, invalid point cloud, clustering failure, etc., and added a global exception capture mechanism to ensure that the failure of single frame point cloud processing will not lead to process interruption or system crash.

[0191] As another embodiment of the present invention, a radar point cloud dynamic and static target differentiation detection device 100 is provided to implement the radar point cloud dynamic and static target differentiation detection method described above, wherein, as Figure 6 As shown, it includes:

[0192] Module 110 is used to acquire radar point cloud data;

[0193] The format conversion module 120 is used to perform adaptive format conversion processing on the radar point cloud data to obtain the format-converted radar point cloud data.

[0194] The point cloud filtering module 130 is used to perform ground point cloud filtering processing on the radar point cloud data after format conversion to obtain non-ground effective target point cloud data.

[0195] The segmentation module 140 is used to perform dynamic and static target segmentation on the non-ground effective target point cloud data according to a multi-dimensional dynamic and static segmentation algorithm including adaptive compensation for velocity error, so as to obtain dynamic point cloud data and static point cloud data.

[0196] The clustering processing module 150 is used to mine the point cloud distribution features of dynamic point cloud data and static point cloud data respectively, and to perform adaptive differential clustering processing on the mined dynamic point cloud distribution features and static point cloud distribution features to obtain dynamic target clustering results and static target clustering results.

[0197] The radar point cloud dynamic and static target differentiation detection device provided by this invention acquires radar point cloud data, performs adaptive format conversion on the radar point cloud data, performs ground point cloud filtering on the converted radar point cloud data, performs dynamic and static target segmentation on non-ground effective target point cloud data according to a multi-dimensional dynamic and static segmentation algorithm, and mines point cloud distribution features for both dynamic and static point cloud data. Then, it performs adaptive differential clustering processing on the mined dynamic and static point cloud distribution features to obtain dynamic target clustering results and static target clustering results. This radar point cloud dynamic and static target differentiation detection device can effectively compensate for the inherent error of radar velocity measurement through a multi-dimensional accurate segmentation algorithm, solving the defect of misjudgment of ambiguous areas caused by traditional fixed threshold segmentation. In addition, through quantitative analysis of dynamic point cloud dispersion and static point cloud clustering, it achieves online adaptive adjustment of clustering parameters, avoiding the essential defect of traditional fixed parameters being unable to match point cloud distribution differences, and effectively improving target detection accuracy. Therefore, the radar point cloud dynamic and static target differentiation detection device provided by the present invention can effectively solve the problems of segmentation error caused by coarse logic in existing algorithms and low detection accuracy caused by low clustering accuracy, thereby improving the target detection accuracy.

[0198] The specific working principle of the radar point cloud dynamic and static target differentiation detection device provided by the present invention can be referred to the description of the radar point cloud dynamic and static target differentiation detection method above, and will not be repeated here.

[0199] As another embodiment of the present invention, a radar point cloud target detection system 10 based on ROS is provided, wherein, as Figure 7 As shown, it includes: a radar sensing device 200 and the radar point cloud dynamic and static target differentiation detection device 100 mentioned above, wherein the radar sensing device 200 is communicatively connected to the radar point cloud dynamic and static target differentiation detection device 100.

[0200] The radar sensing device 200 is used to collect radar point cloud data;

[0201] The radar point cloud dynamic and static target differentiation detection device 100 is used to perform dynamic and static differentiation processing on the radar point cloud data to obtain dynamic target clustering results and static target clustering results.

[0202] In this embodiment of the invention, the ROS-based radar point cloud target detection system 10 is specifically a hardware and software combined architecture. The ROS architecture can be deployed on an industrial control computer or embedded development board running the Ubuntu operating system. The industrial control computer or embedded development board can include the radar point cloud dynamic and static target differentiation detection device 100. The industrial control computer or development board establishes a communication connection with a radar sensor (millimeter-wave radar / liquid radar) that supports outputting four-dimensional information of x, y, z, and velocity through Ethernet, CAN bus, etc.

[0203] The ROS-based radar point cloud target detection system can specifically achieve target detection function through the radar point cloud dynamic and static target differentiation detection device 100.

[0204] In this embodiment of the invention, to facilitate the implementation of the radar point cloud dynamic and static target differentiation detection device 100, the ROS-based radar point cloud target detection system 10 also includes a ROS node initialization and parameter configuration module. This ROS node initialization and parameter configuration module specifically starts a radar detection ROS node named "radar_detection", reads all core parameters configured externally by the launch file through the ROS private namespace, and completes parameter verification and log output. At the same time, it initializes the point cloud subscription communication link, the clustering result publishing communication link, and the debug point cloud publishing communication link. The size of all communication link queues can be adjusted through external configuration, and is set to 10 by default to ensure the real-time transmission and processing of high-frequency radar point cloud data of 10Hz and above.

[0205] Specifically, the core parameters include: input topic name, output topic name, dynamic and static velocity distinction benchmark threshold V_base, velocity error compensation coefficient k_v, dynamic clustering base eps value eps_d0, dynamic clustering base min_samples value min_s_d0, static clustering base eps value eps_s0, static clustering base min_samples value min_s_s0, ground removal z-coordinate threshold interval (z_min, z_max), point cloud density sampling interval d_sample, clustering parameter self-calibration iteration count N_iter, and multi-frame temporal verification window length T. Among these, the ratio of static clustering base eps value to dynamic clustering base eps value can be adjusted through external configuration, with a recommended configuration range of 1.2 to 2 times. This range represents the optimal initial ratio range derived based on the distribution characteristics of dynamic and static point clouds.

[0206] In summary, the ROS-based radar point cloud target detection system provided by this invention effectively compensates for the inherent errors in radar velocity measurement through a multi-dimensional precise segmentation algorithm, solving the defect of misjudgment of ambiguous regions caused by traditional fixed threshold segmentation. Furthermore, through quantitative analysis of dynamic point cloud dispersion and static point cloud clustering, online adaptive adjustment of clustering parameters is achieved, avoiding the inherent defect of traditional fixed parameters failing to match differences in point cloud distribution, thus effectively improving target detection accuracy. Moreover, this ROS-based radar point cloud target detection system is developed based on the open-source ROS architecture, compatible with both ROS 1 and ROS 2 architectures. All dependencies are open-source general-purpose packages, requiring no additional hardware modifications; all functions can be implemented simply through software module integration. The system can be deployed on various hardware platforms such as conventional industrial control computers, embedded development boards, and vehicle domain controllers, resulting in low development and deployment costs and easy large-scale application in engineering projects.

[0207] The specific working principle of the radar point cloud dynamic and static target differentiation detection system provided by this invention can be referred to the description of the radar point cloud dynamic and static target differentiation detection method above, and will not be repeated here.

[0208] It is understood that the above embodiments are merely exemplary implementations used to illustrate the principles of the present invention, and the present invention is not limited thereto. For those skilled in the art, various modifications and improvements can be made without departing from the spirit and essence of the present invention, and these modifications and improvements are also considered to be within the scope of protection of the present invention.

Claims

1. A method for differential detection of dynamic and static targets in radar point clouds, characterized in that, include: Acquire radar point cloud data; The radar point cloud data is subjected to adaptive format conversion processing to obtain radar point cloud data after format conversion. The radar point cloud data after format conversion is filtered by ground point cloud to obtain non-ground effective target point cloud data. The non-ground effective target point cloud data is segmented into dynamic and static targets using a multi-dimensional dynamic and static segmentation algorithm, including adaptive compensation for velocity error, to obtain dynamic point cloud data and static point cloud data. Point cloud distribution features are mined from both dynamic and static point cloud data. Adaptive differential clustering is then applied to the mined dynamic and static point cloud distribution features to obtain dynamic and static target clustering results.

2. The radar point cloud dynamic and static target differentiation detection method according to claim 1, characterized in that, The non-ground effective target point cloud data is segmented into dynamic and static targets using a multi-dimensional dynamic and static segmentation algorithm, including adaptive velocity error compensation, to obtain dynamic point cloud data and static point cloud data, including: The reference speed threshold is dynamically adjusted according to the preset speed error compensation coefficient to obtain the dynamic speed threshold and the static speed threshold. The non-ground effective target point cloud data is initially segmented based on the dynamic velocity threshold and the static velocity threshold to obtain preliminary segmentation results. The preliminary segmentation results include candidate dynamic point clouds, candidate static point clouds, and fuzzy region point clouds. The candidate dynamic point clouds are the set of point clouds in the non-ground effective target point cloud data whose absolute point cloud velocity is greater than the dynamic velocity threshold. The candidate static point clouds are the set of point clouds in the non-ground effective target point cloud data whose absolute point cloud velocity is less than the static velocity threshold. The fuzzy region point clouds are the set of point clouds in the non-ground effective target point cloud data whose absolute point cloud velocity is greater than or equal to the static velocity threshold and less than or equal to the dynamic velocity threshold. A secondary discrimination is performed on the fuzzy region point cloud in the preliminary segmentation result to obtain the secondary discrimination result of the fuzzy region point cloud. The preliminary segmentation results and the secondary discrimination results of the fuzzy region point cloud are fused to obtain a dynamic point cloud set and a static point cloud set.

3. The radar point cloud dynamic and static target differentiation detection method according to claim 2, characterized in that, The point cloud of the fuzzy region in the preliminary segmentation result is subjected to secondary discrimination to obtain the secondary discrimination result of the fuzzy region point cloud, including: For each point cloud in the fuzzy region point cloud, calculate the number of point clouds in the neighborhood centered on that point cloud and with a preset radius. If the number of point clouds in the neighborhood is less than the preset neighborhood density threshold, then the point cloud in the fuzzy region point cloud is determined as a candidate dynamic point cloud. If the number of point clouds in the neighborhood is not less than the preset neighborhood density threshold, then determine whether the coordinate offset of the point cloud in the blurred region point cloud within a consecutive preset length of time frame is greater than the preset offset threshold. If the offset exceeds the preset threshold, then the point cloud in the blurred region point cloud is determined as a candidate dynamic point cloud; If the offset is not greater than the preset offset threshold, then the point cloud in the blurred region point cloud is determined as a candidate dynamic point cloud or a candidate static point cloud.

4. The radar point cloud dynamic and static target differentiation detection method according to claim 1, characterized in that, Point cloud distribution features are mined from both dynamic and static point cloud data. Adaptive differential clustering is then applied to the mined dynamic and static point cloud distribution features to obtain dynamic and static target clustering results, including: Point cloud distribution features are quantized for dynamic point cloud data and static point cloud data respectively to obtain dynamic point cloud quantization features and static point cloud quantization features. The dynamic clustering parameters are iteratively calibrated based on the dynamic point cloud quantization features to obtain calibrated dynamic clustering parameters, and the static clustering parameters are iteratively calibrated based on the static point cloud quantization features to obtain calibrated static clustering parameters. The dynamic initial effective clustering results are obtained by density clustering based on the x and y plane coordinates of the point cloud according to the calibrated dynamic clustering parameters, and the static initial effective clustering results are obtained by density clustering based on the x and y plane coordinates of the point cloud according to the calibrated static clustering parameters. The dynamic initial valid clustering results and the static initial valid clustering results are filtered and the target attributes are verified twice to obtain the dynamic target clustering results and the static target clustering results.

5. The radar point cloud dynamic and static target differentiation detection method according to claim 4, characterized in that, Point cloud distribution features are quantized for both dynamic and static point cloud data to obtain dynamic point cloud quantization features and static point cloud quantization features, including: The point cloud x and y planes in the dynamic point cloud data are divided according to the preset basic sampling interval to obtain multiple sampling grids; Calculate the point cloud density within each sampling grid; The dynamic point cloud discreteness feature value is determined based on the standard deviation of the point cloud density in all sampling grids, and the dynamic point cloud discreteness feature value is determined as the dynamic point cloud quantization feature. For each point in the static point cloud data, calculate its average distance to K nearest neighbors, where K = min_s_s0, min_s_s0 represents the basic min_samples value of static clustering, and min_samples represents the minimum number of data points required for an effective cluster; The clustering feature value is obtained by taking the mean of the average distances corresponding to all points in the static point cloud data and then calculating the reciprocal of the mean. The aggregation feature value is determined as the static point cloud quantization feature.

6. The radar point cloud dynamic and static target differentiation detection method according to claim 5, characterized in that, The process includes iteratively calibrating dynamic clustering parameters based on the dynamic point cloud quantization features to obtain calibrated dynamic clustering parameters, and iteratively calibrating static clustering parameters based on the static point cloud quantization features to obtain calibrated static clustering parameters, including: Dynamic clustering parameters are determined based on the dynamic point cloud quantization features and the preset dynamic point cloud standard discreteness feature values, and the dynamic clustering parameters are calibrated. The expression for dynamic clustering parameter calibration is as follows: eps_d = eps_d0 × (1 + 0.5×(ρ_d0 / ρ_d - 1)), min_samples_d = max(round(min_s_d0 × (ρ_d0 / ρ_d)), 1), Where eps_d represents the dynamic clustering parameters, eps_d0 represents the basic eps value of dynamic clustering, ρ_d0 represents the preset standard discreteness feature value of dynamic point cloud, ρ_d represents the discreteness feature value of dynamic point cloud, min_samples_d represents the minimum number of data points required for a valid cluster in dynamic clustering, min_s_d0 represents the basic min_samples value of dynamic clustering, and eps represents the neighborhood search radius of the clustering algorithm; Static clustering parameters are determined based on the static point cloud quantization features and preset static point cloud standard clustering feature values, and the static clustering parameters are calibrated. The expression for static clustering parameter calibration is as follows: eps_s = eps_s0 × (1 - 0.3×(α_s / α_s0 - 1)), min_samples_s = max(round(min_s_s0 × (α_s / α_s0)), 3), Where eps_s represents the static clustering parameters, eps_s0 represents the basic eps value of static clustering, α_s0 represents the preset standard clustering feature value of static point cloud, α_s represents the static point cloud quantization feature, min_samples_s represents the minimum number of data points required for a valid cluster in static clustering, and min_s_s0 represents the basic min_samples value of static clustering. The calibration process for both dynamic and static clustering parameters is iterated a preset number of times, and the clustering effect evaluation index is calculated after each iteration. If the difference between the clustering performance evaluation index of two consecutive iterations is less than the preset convergence threshold, the iteration is determined to be converged and the iteration is stopped. The dynamic clustering parameters and static clustering parameters are output. If the iteration fails to converge after a preset number of iterations, the calibration results corresponding to the dynamic clustering basic parameters and the static clustering basic parameters will be output respectively.

7. The radar point cloud dynamic and static target differentiation detection method according to any one of claims 1 to 6, characterized in that, The radar point cloud data undergoes adaptive format conversion to obtain format-converted radar point cloud data, including: The radar point cloud data is automatically adapted to the point cloud field structure, and the four-dimensional valid fields after the field structure adaptation are read. Filter outliers in the velocity field among the valid fields in the four-dimensional array; Convert the four-dimensional valid fields after outlier filtering into a preset unified data format.

8. The radar point cloud dynamic and static target differentiation detection method according to any one of claims 1 to 6, characterized in that, The converted radar point cloud data undergoes ground point cloud filtering to obtain non-ground effective target point cloud data, including: The radar point cloud data after format conversion is subjected to preliminary ground point cloud filtering processing according to the preset ground point cloud filtering threshold range to obtain non-ground candidate point clouds. Normal vector estimation is performed on the non-ground candidate point cloud to perform secondary filtering of the ground point cloud based on the angle between the normal vector of each point cloud in the non-ground candidate point cloud and the ground normal vector, so as to obtain the non-ground effective target point cloud data.

9. A radar point cloud dynamic and static target differentiation detection device, used to implement the radar point cloud dynamic and static target differentiation detection method according to any one of claims 1 to 8, characterized in that, include: The acquisition module is used to acquire radar point cloud data; The format conversion module is used to perform adaptive format conversion processing on the radar point cloud data to obtain the format-converted radar point cloud data. The point cloud filtering module is used to perform ground point cloud filtering on the radar point cloud data after format conversion to obtain non-ground effective target point cloud data. The segmentation module is used to perform dynamic and static target segmentation on the non-ground effective target point cloud data according to a multi-dimensional dynamic and static segmentation algorithm including adaptive compensation for velocity error, so as to obtain dynamic point cloud data and static point cloud data. The clustering module is used to mine point cloud distribution features of dynamic point cloud data and static point cloud data respectively, and to perform adaptive differential clustering processing on the mined dynamic point cloud distribution features and static point cloud distribution features to obtain dynamic target clustering results and static target clustering results.

10. A radar point cloud target detection system based on ROS, characterized in that, include: The radar sensing device and the radar point cloud dynamic and static target differentiation detection device as described in claim 9, wherein the radar sensing device is communicatively connected to the radar point cloud dynamic and static target differentiation detection device. The radar sensing device is used to collect radar point cloud data; The radar point cloud dynamic and static target differentiation detection device is used to perform dynamic and static differentiation processing on the radar point cloud data to obtain dynamic target clustering results and static target clustering results.

Images