A target heading angle estimation method, device, equipment and medium of a vehicle-mounted radar

Abstract

The application provides a target heading angle estimation method, device and equipment of a vehicle-mounted radar and a medium. It relates to the technical field of vehicle-mounted radar perception. The method comprises: clustering the original point cloud data of the continuous frames of the vehicle-mounted radar to obtain at least one target point cloud cluster; generating multiple candidate heading angle hypotheses for each target point cloud cluster in the initial frame, scoring the multiple candidate heading angle hypotheses of each target point cloud cluster in the non-initial frame, screening the multiple candidate heading angle hypotheses based on the score to obtain the effective heading angle hypothesis of each target point cloud cluster; extracting the features of each target point cloud cluster to obtain the feature vector of each target point cloud cluster; for each effective heading angle hypothesis, based on the feature vector of the corresponding target point cloud cluster; inputting the feature vector into a pre-trained estimation network to map the features of the target point cloud cluster to the heading angle prediction value and the uncertainty of the corresponding target point cloud cluster.

Description

Technical Field

[0001] This invention relates to the field of vehicle-mounted radar sensing technology, and in particular to a method, apparatus, device, and medium for estimating the target heading angle of vehicle-mounted radar. Background Technology

[0002] In the field of vehicle-mounted radar perception technology, accurately estimating the heading angle of dynamic targets (such as vehicles and pedestrians) is a crucial step in environmental perception. The estimation result directly impacts the performance of subsequent modules such as target trajectory prediction, behavioral intent judgment, and path planning. Existing technologies for estimating target heading angles based on millimeter-wave radar point clouds mainly include principal component analysis (PCA) and minimum bounding rectangle fitting. Specifically, this method calculates the covariance matrix of the target point cloud cluster to obtain its eigenvectors to determine the principal axis direction, or fits the minimum bounding rectangle of the point cloud using a rotating caliper algorithm, and uses the long side of the rectangle as the estimated heading angle of the target. This technique relies entirely on the geometric distribution of the point cloud, assuming that the point cloud can completely and uniformly reflect the target's outline.

[0003] However, the accuracy of principal component analysis (PCA) and minimum bounding rectangle fitting for point clouds is highly dependent on the density, uniformity of distribution, and shape integrity of the point cloud. In actual automotive radar operating scenarios, point clouds generated by distant targets, low-reflectivity targets, or partially occluded targets often exhibit sparse characteristics, with unclear or incomplete contour edges. This core mechanism makes the principal axis direction or long side direction of the rectangle obtained by geometric fitting highly susceptible to significant deviations due to a few outliers or sparse regions, resulting in high uncertainty and drastic inter-frame jumps in the heading angle estimation results. The accuracy and robustness of the heading angle estimation are severely insufficient, failing to meet the perception stability requirements of autonomous driving systems. Summary of the Invention

[0004] Therefore, it is necessary to provide a method, device, equipment, and medium for estimating the target heading angle of vehicle-mounted radar to address the aforementioned technical problems.

[0005] The following technical solution is adopted in this specification: This manual provides a method for estimating the target heading angle of a vehicle-mounted radar, including: Acquire raw point cloud data of consecutive frames from the vehicle-mounted radar; Clustering is performed on the raw point cloud data to obtain at least one target point cloud cluster; Within the initial frame, multiple candidate heading angle hypotheses are generated for each target point cloud cluster. For each non-initial frame, based on the degree of fit between each candidate heading angle hypothesis and the geometry of the target point cloud cluster, the consistency with the direction of the centroid movement vector of the target point cloud cluster, and the consistency with the reference direction of the driving environment, the multiple candidate heading angle hypotheses for each target point cloud cluster are scored independently. Based on the scoring results, the multiple candidate heading angle hypotheses for each target point cloud cluster are independently filtered to obtain at least one valid heading angle hypothesis for each target point cloud cluster. For each target point cloud cluster, feature extraction is performed to obtain the feature vector of each target point cloud cluster; for each effective heading angle hypothesis of each target point cloud cluster, the feature vector of the corresponding target point cloud cluster is input into the pre-trained estimation network, and the estimation network maps the feature vector to the heading angle prediction value and uncertainty of the corresponding effective heading angle hypothesis; For multiple valid heading angle assumptions for the same target point cloud cluster, the heading angle prediction with the smallest uncertainty is selected as the final heading angle estimation result of the target point cloud cluster in the current frame.

[0006] Furthermore, before performing clustering processing on the original point cloud data, preprocessing of the original point cloud data is included, specifically including: Based on the preset radar detection distance threshold and the effective detection range of the road, the raw point cloud data that exceeds the distance constraint is filtered out; Filter out raw point cloud data with radar reflection intensity below the radar reflection intensity threshold; Filter out raw point cloud data whose velocity amplitude is lower than the velocity amplitude threshold and whose velocity variance is higher than the preset fluctuation threshold; Static clutter point clouds generated by ground reflectors or fixed structures can be filtered out by employing static background subtraction algorithms or ground plane filtering algorithms.

[0007] Furthermore, the independent scoring of multiple candidate heading angle assumptions for each target point cloud cluster specifically includes: For each target point cloud cluster, the principal axis direction of the target point cloud cluster is determined by principal component analysis, and the deviation between each candidate heading angle hypothesis of the target point cloud cluster and the principal axis direction is used as the geometric fit score of the candidate heading angle hypothesis. Obtain the centroid movement vector of each target point cloud cluster, and use the deviation between each candidate heading angle hypothesis and the centroid movement vector as the velocity direction consistency score of the candidate heading angle hypothesis; The deviation between each candidate heading angle assumption and the driving environment reference direction is used as the environmental reference consistency score of the candidate heading angle assumption; Based on the preset scoring weights, the geometric fitting score, velocity direction consistency score, and environmental reference consistency score are weighted and summed to obtain the frame score for each candidate heading angle hypothesis.

[0008] Furthermore, the independent screening of multiple candidate heading angle hypotheses for each target point cloud cluster based on the scoring results specifically includes: Set a preset number of sliding window frames, and for each candidate heading angle hypothesis of each target point cloud cluster, accumulate the frame scores of the candidate heading angle hypothesis in all frames within the sliding window to obtain the cumulative score of the candidate heading angle hypothesis; Determine whether the cumulative score of each candidate heading angle hypothesis is lower than a preset elimination threshold, or determine whether the candidate heading angle hypothesis is the lowest score within the window within a preset number of consecutive frames; If any of the above conditions are met, the candidate heading angle assumptions are eliminated; the remaining candidate heading angle assumptions in each target point cloud cluster are retained as the effective heading angle assumptions for the target point cloud cluster.

[0009] Furthermore, the step of extracting features from each target point cloud cluster to obtain the feature vector of each target point cloud cluster specifically includes: PCA analysis was performed on the preprocessed target point cloud cluster to obtain the principal component directions of the major and minor axes. At the same time, a rotation bounding rectangle was fitted to the target point cloud cluster and the side directions of the rotation bounding rectangle were obtained. Extract the geometric features of the target point cloud cluster, including the directional distribution of the normals of the edge points of the target point cloud cluster, aspect ratio, left-right symmetry and front-back symmetry of the graphic; Statistical histograms of radial velocity distribution are plotted for each component of the target point cloud cluster, and micro-Doppler features are extracted. The micro-Doppler features include the maximum energy range of the histogram, the velocity distribution width at the energy concentration point, the energy ratio of the positive and negative velocity ranges, the velocity distribution skewness, and the kurtosis. Extract the reflection intensity features of each component of the target point cloud cluster. The reflection intensity features include the average reflection intensity, the maximum reflection intensity, and the proportion of strong reflection points that are stronger than a preset threshold. Extract the positional features of strong reflection points, including the deviation of strong reflection points from the boundary of the rotating enclosing rectangle, the deviation of strong reflection points from the geometric center of the target cloud cluster, and the orientation concentration features of strong reflection points. After normalizing the geometric features, micro-Doppler features, reflection intensity features, and strong reflection point location features, the feature vectors of each target point cloud cluster are obtained by concatenating them.

[0010] Further, the step of inputting the feature vector into a pre-trained estimation network, and having the estimation network map the feature vector to the predicted heading angle and uncertainty corresponding to the effective heading angle assumption, specifically includes: For each effective heading angle assumption of each target point cloud cluster, the feature vector of the target point cloud cluster in the current frame, the centroid movement vector of the past preset number of frames, and the driving environment reference direction are used as inputs to the pre-trained estimation network. The estimation network first extracts features and adds attention weights to the feature vectors through the geometric feature branch, and then performs temporal encoding on the centroid movement vector time series and the driving environment reference direction through the motion feature branch; The output features of the geometric feature branch and the motion feature branch are concatenated and fused. After mapping through a fully connected layer, the predicted heading angle corresponding to the effective heading angle assumption is output. At the same time, the uncertainty index, which characterizes the accuracy of the predicted value estimation, is output. The uncertainty index is the standard deviation of the heading angle estimate.

[0011] Furthermore, the pre-trained estimation network is a dual-branch fusion neural network that incorporates an error monitoring mechanism and a self-correction mechanism; The dual-branch fusion neural network includes a geometric component feature branch and a velocity environment reference branch. The geometric component feature branch consists of a fully connected layer, a normalization layer, an activation layer, and an attention weight module, which is used to extract the feature vector of the target point cloud cluster by weight. The velocity environment reference branch consists of a GRU or LSTM network, which is used to encode the temporal feature of the centroid movement vector time series and the driving environment reference direction. The error monitoring mechanism monitors the heading angle estimation uncertainty and geometric direction deviation by setting a threshold. If the uncertainty exceeds the uncertainty threshold for multiple consecutive frames, or if the deviation between the heading angle prediction and the geometric fitting direction exceeds the geometric direction deviation threshold, the heading angle estimation is marked as a low confidence estimation. The self-correction mechanism includes extrinsic yaw correction, component feature weight adjustment, and environmental reference auxiliary correction. Extrinsic yaw correction is used to estimate and eliminate fixed yaw errors caused by radar installation. Component feature weight adjustment is used to reduce the feature weight of components with unstable features and low point cloud density. Environmental reference auxiliary correction is used to increase the weight of the driving environment reference direction in network fusion when the driving environment reference direction is reliable.

[0012] This specification provides a target heading angle estimation device for vehicle-mounted radar, including: The data acquisition module is used to acquire raw point cloud data of continuous frames from the vehicle-mounted radar. The target point cloud cluster acquisition module is used to perform clustering processing on the raw point cloud data to obtain at least one target point cloud cluster. The effective heading angle hypothesis acquisition module is used to generate multiple candidate heading angle hypotheses for each target point cloud cluster in the initial frame; for each non-initial frame, based on the degree of fit between each candidate heading angle hypothesis and the geometry of the target point cloud cluster, the consistency with the direction of the centroid movement vector of the target point cloud cluster, and the consistency with the reference direction of the driving environment, the multiple candidate heading angle hypotheses of each target point cloud cluster are scored independently, and based on the scoring results, the multiple candidate heading angle hypotheses of each target point cloud cluster are independently filtered to obtain at least one effective heading angle hypothesis for each target point cloud cluster; The prediction module is used to extract features from each target point cloud cluster to obtain the feature vector of each target point cloud cluster. For each effective heading angle hypothesis of each target point cloud cluster, the feature vector of the corresponding target point cloud cluster is input into the pre-trained estimation network, and the estimation network maps the feature vector to the heading angle prediction value and uncertainty of the corresponding effective heading angle hypothesis. The target heading angle estimation module is used to select the heading angle prediction value with the smallest uncertainty from multiple valid heading angle assumptions for the same target point cloud cluster as the final heading angle estimation result of the target point cloud cluster in the current frame.

[0013] This specification provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the target heading angle estimation method for vehicle-mounted radar described above.

[0014] This specification provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the target heading angle estimation method of the vehicle-mounted radar described above.

[0015] The above-mentioned technical solutions adopted in this specification can achieve the following beneficial effects: This invention generates multiple candidate heading angle hypotheses independently for each target point cloud cluster in the initial frame. In each subsequent frame, based on the degree of fit between each hypothesis and its corresponding point cloud cluster's geometry, consistency of the centroid movement vector direction, and consistency of the driving environment reference direction, multiple hypotheses within the same target point cloud cluster are independently weighted and accumulated through a sliding window. Hypotheses with low scores or consistently the lowest scores are eliminated, ensuring that each target point cloud cluster retains only at least one valid hypothesis. This eliminates ambiguous headings introduced by sparse point clouds, occlusion, or ambiguous velocity directions at an early stage, preventing invalid hypotheses from propagating to subsequent estimation networks. Based on this, for each retained valid hypothesis, the feature vector of the corresponding target point cloud cluster is extracted and input into a pre-trained estimation network. This network simultaneously outputs the predicted heading angle value under that hypothesis and the uncertainty representing the estimation accuracy. Since multiple valid hypotheses may exist for the same target point cloud cluster (such as the ambiguity of the orientation of a stationary target), this invention further compares the uncertainties corresponding to multiple hypotheses for the same target point cloud cluster and selects the predicted heading angle value with the lowest uncertainty as the final estimation result for the current frame.

[0016] The aforementioned multi-criteria scoring and dynamic elimination mechanism utilizes geometry to compress the hypothesis space with low computational cost, concentrating network input on high-confidence candidate directions. Meanwhile, the uncertainty of the estimated network output provides a quantitative basis for the final selection, enabling the most reliable heading angle to be automatically selected without external rules in ambiguous scenarios. The combination of these two mechanisms achieves full-process optimization of heading angle estimation results from hypothesis generation to value judgment under complex conditions such as incomplete point clouds, low-speed or stationary targets, and available reference parts of the driving environment, thereby improving the stability and quantifiable confidence of the estimation. Attached Figure Description

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

[0018] Figure 1 A flowchart illustrating a target heading angle estimation method for vehicle-mounted radar provided in this specification; Figure 2 This is a schematic diagram of the target heading angle estimation device for vehicle-mounted radar provided in this specification; Figure 3 This is a schematic diagram of a computer device provided for this specification. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this specification clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments in this specification without creative effort are within the scope of protection of this application.

[0020] This invention belongs to the field of vehicle-mounted radar perception technology, specifically relating to a vehicle-mounted millimeter-wave radar target heading angle estimation system, particularly a heading angle estimation system that combines location-level micro-Doppler features, geometric features, time series networks, uncertainty assessment, and environmental reference correction.

[0021] With the development of autonomous driving and intelligent assisted driving systems, accurate perception of the dynamic state of surrounding targets (especially heading / orientation / azimuth angle) is becoming increasingly important. The heading angle is a key input for multiple modules, including predicting the target's future trajectory, determining its movement intention, path planning, and collision prevention. However, current technologies using millimeter-wave radar point clouds to estimate target heading angles have several limitations and challenges, mainly including the following: Traditional methods typically employ principal component analysis (PCA) of the point cloud or a minimum bounding box to fit the target's geometric major / minor axis directions to infer the heading angle. These methods heavily rely on the density, uniformity of distribution, and shape integrity of the point cloud. At long distances, with small targets, low reflectivity targets, or in partially obscured conditions, the point cloud is sparse, and the contour edges are unclear or incomplete, leading to large errors in estimating the major and minor axis directions of the geometric fit. Furthermore, estimating the heading angle using the target's velocity vector or historical trajectory direction is one of the commonly used methods. This method works well when the target's movement is stable, its speed is sufficiently high, and its trajectory is continuous and clear. However, for stationary or slowly moving targets, velocity direction information is weak or unavailable; or there is a deviation between the velocity direction and orientation (e.g., the target may be sideways or turning), leading to distorted heading estimation. Millimeter-wave radar can provide radial velocity information (Doppler component), which theoretically can provide clues about the orientation of certain parts of the target. In existing methods, the Doppler signal is often used as a whole or only as part of the velocity, without joint analysis with geometric edge / component features to improve the stability of heading estimation for stationary or slowly moving targets.

[0022] Therefore, there is an urgent need for a new heading angle estimation method in existing technologies. This patented technology aims to provide an innovative solution to this key problem. This invention provides a vehicle-mounted millimeter-wave radar target heading angle estimation system that maintains high accuracy and reliability under various complex scenarios, including static and dynamic targets, sparse and complete point clouds, strong and weak velocity / trajectory information, and available and unavailable environmental references. It also provides estimation confidence and error monitoring and self-correction capabilities to improve the stability and reliability of heading angle estimation in autonomous driving systems across all scenarios. The vehicle-mounted millimeter-wave radar target heading angle estimation method described in this invention maintains low error even with stationary or moving targets, sparse / noisy point clouds, and incomplete reflection conditions, and has confidence output and self-correction capabilities. This method includes several main modules: data preprocessing, feature extraction, hypothesis generation and dynamic elimination, time-series fusion estimation, and uncertainty assessment and correction.

[0023] The target heading angle estimation method of the vehicle-mounted radar of the present invention is described below with reference to the accompanying drawings.

[0024] Figure 1This is a flowchart illustrating a target heading angle estimation method for vehicle-mounted radar provided in this specification, as shown below. Figure 1 As shown, the method includes the following: S101. Obtain raw point cloud data of continuous frames from the vehicle radar.

[0025] For example, in this embodiment, the vehicle-mounted millimeter-wave radar continuously scans the detection road area, and performs signal processing operations such as signal calculation and filtering on each frame of radar echo to obtain raw point cloud data containing multi-dimensional feature information such as distance, azimuth angle, elevation angle, radial velocity / Doppler response, and point cloud reflection intensity.

[0026] Because the original point cloud data contains outliers such as noise clutter and false targets that are clearly inconsistent with the target's movement trend, targeted point cloud cleaning processing is required. Taking the radar installation location as the coordinate origin and combining the actual road conditions, the detection and research area is defined as a rectangular area. By constraining the longitudinal and lateral detection range of the radar target, valid target point clouds are screened. First, invalid point cloud data that exceeds the detection road range and the radar's maximum detection distance requirement are removed. Then, a reflection intensity threshold RCS_min is set and applied to remove point cloud data with reflection intensity lower than RCS_min. At the same time, point cloud data with velocity amplitude lower than the preset velocity threshold and velocity values ​​that fluctuate greatly with noise are also removed. Finally, a static background subtraction algorithm or a ground plane filtering algorithm is used to filter out static clutter point cloud data generated by ground reflection, bridge piers, guardrails and other fixed road structures, completing the entire process of point cloud denoising.

[0027] After point cloud denoising is completed, the centroid position of each target in each frame is determined and recorded based on the radar coordinate system. At the same time, the centroid position of adjacent frames is compared to calculate the inter-frame position change. If the centroid movement reaches a preset threshold, the centroid movement vector of the target is calculated and used as input data for velocity and trajectory history in subsequent heading angle estimation. In addition, to address the problem that newly detected targets are prone to centroid position jumps, a sliding filter algorithm is used to perform temporal smoothing on the centroid position of targets in each frame, effectively reducing centroid position fluctuations caused by noise. Finally, preprocessed point cloud data with denoising and smoothing is obtained, providing a reliable data foundation for subsequent target point cloud clustering, feature extraction and other operations.

[0028] S102. Perform clustering processing on the original point cloud data to obtain at least one target point cloud cluster.

[0029] For example, a density-based clustering algorithm (such as DBSCAN) is used to cluster the preprocessed point cloud in each frame. A clustering radius and a minimum point count threshold are set to group spatially adjacent and density-connected points into the same target point cloud cluster. After clustering, each target point cloud cluster corresponds to a physical target (such as a vehicle or pedestrian), and the contour point set, centroid coordinates, and radial velocity distribution of each cluster are labeled. For the association of the same target in consecutive frames, nearest neighbor matching or the Hungarian algorithm is used for inter-frame data association to maintain the target's unique temporal identifier.

[0030] S103. In the initial frame, generate multiple candidate heading angle hypotheses for each target point cloud cluster. For each non-initial frame, based on the degree of fit between each candidate heading angle hypothesis and the geometry of the target point cloud cluster, the consistency with the direction of the centroid movement vector of the target point cloud cluster, and the consistency with the reference direction of the driving environment, independently score the multiple candidate heading angle hypotheses for each target point cloud cluster. Based on the scoring results, independently filter the multiple candidate heading angle hypotheses for each target point cloud cluster to obtain at least one valid heading angle hypothesis for each target point cloud cluster.

[0031] For example, within the initial frame of target detection, K uniformly distributed candidate heading angle hypotheses are generated for each target point cloud cluster. For instance, K=8, corresponding to candidate heading angles of... For each subsequent frame, an assumption is made for each candidate heading angle of each target point cloud cluster. Three scoring indicators were calculated separately. The geometric error was calculated with reference to the direction of the long side obtained by fitting the target point cloud cluster with a rotated bounding rectangle. The absolute deviation between the candidate heading angle and the reference direction was also calculated. The smaller the deviation, the higher the geometric fit score; the consistency of velocity direction is based on the inter-frame centroid movement vector of the target point cloud cluster, when the target's velocity... When the speed exceeds a preset speed threshold v_thresh (e.g., 2 m / s), calculate the absolute deviation between the candidate heading angle and the direction of centroid movement θ_vel. The smaller the deviation, the higher the corresponding speed-direction consistency score. If the target's moving speed is less than v_thresh, the weight of this score is set to zero or a minimum value. Environmental reference consistency uses lane lines, road edges, visual signs, or high-precision maps to provide driving environment reference directions. Using this as a reference, calculate the absolute deviation between the candidate heading angle and the reference direction. The smaller the deviation, the higher the environmental reference consistency score. The weight of this score is dynamically set according to the reliability of the reference direction (such as map accuracy and lane detection confidence). Based on the preset score weights, the scores corresponding to the three indicators are weighted and summed to obtain the frame score of each candidate heading angle hypothesis in the current frame, which provides a basis for subsequent hypothesis screening.

[0032] For example, for a dynamic elimination rule based on the candidate heading angle assumption, the number of sliding window frames is first defined. For each candidate heading angle assumption i of each target point cloud cluster, in the past Intra-frame cumulative calculation of overall score The calculation formula is: ; in, , , The deviation mapping scoring function can be implemented using forms such as the Gaussian decay function, mapping geometric deviation, velocity direction deviation, and environmental reference deviation to corresponding scores. The smaller the deviation, the higher the mapping score. , , The preset weights for the three indicators—geometric fit, velocity consistency, and environmental reference consistency—are then defined. Subsequently, elimination criteria are set, based on the cumulative score of a candidate heading angle hypothesis i. For at least L consecutive frames (e.g., L=3), the score must be the lowest within the corresponding sliding window, or If the candidate heading angle hypothesis is lower than the preset low score threshold S_thresh_low, it will be eliminated. Finally, the top P candidate heading angle hypotheses with the highest cumulative scores (e.g., P=2) will be retained as valid heading angle hypotheses for each target point cloud cluster, and used in the subsequent network fusion and heading angle estimation stages.

[0033] S104. Extract features from each target point cloud cluster to obtain the feature vector of each target point cloud cluster. For each effective heading angle hypothesis of each target point cloud cluster, input the feature vector of the corresponding target point cloud cluster into the pre-trained estimation network. The estimation network maps the feature vector to the heading angle prediction value and uncertainty of the corresponding effective heading angle hypothesis.

[0034] For example, based on the preprocessed radar target point cloud information, the point cloud within a single frame is first clustered to obtain each target point cloud cluster. Then, feature extraction is performed on each target point cloud cluster. First, PCA analysis is performed on the target point cloud cluster to obtain the two main principal component directions, the major axis and the minor axis. At the same time, a rotational bounding rectangle is fitted to the entire target point cloud cluster to obtain the direction of the long side of the rectangle. Along the direction of the shorter side.

[0035] Subsequently, the geometric features of the target point cloud cluster are extracted. Edge points that are far from the bounding rectangle or deviate significantly from the principal axis are extracted at the edge of the point cloud. Normals are fitted to these edge points and their directional distribution is estimated. At the same time, the geometric features of the target point cloud cluster, such as aspect ratio, left-right symmetry, and front-back symmetry, are statistically analyzed.

[0036] For each component of the target point cloud cluster, a radial velocity distribution histogram is statistically analyzed in a single frame. The preset velocity range is [−V_max, −V_max+Δv, …, V_max] (e.g., V_max=20m / s, Δv=1m / s). From the histogram, the range of maximum energy, the velocity distribution width at the energy concentration point (e.g., the range exceeding 80% of the total energy), the energy ratio of the positive velocity range to the negative velocity range, the velocity distribution skewness, and the micro-Doppler characteristics of the kurtosis are extracted.

[0037] Simultaneously, the reflection intensity characteristics of each component are extracted, and the average RCS_mean, maximum RCS_max, and percentage of strong reflection points stronger than the preset threshold RCS_high are calculated for each component.

[0038] Finally, the positional features of strong reflection points are extracted, and the deviation of strong reflection points from the boundary of the rotating enclosing rectangle and the geometric center of the target point cloud cluster is measured. If there are features of strong reflection points concentrated in a certain direction, they are used as auxiliary clues for the target orientation.

[0039] The extracted geometric features, micro-Doppler features, reflection intensity features, and strong reflection point location features are normalized and then concatenated to obtain the feature vectors of each target point cloud cluster.

[0040] For the P retained candidate heading angle hypotheses, a feature vector is constructed for each hypothesis in each frame. Specifically, this includes the micro-Doppler feature vectors of each target component (covering spectral peak position, energy spectral width, energy ratio between positive and negative velocity ranges, velocity distribution skewness and kurtosis, etc.), target geometric edge direction features (including edge point normal direction distribution, aspect ratio, graphic symmetry, etc.), and target RCS features (including average and maximum reflection intensity, proportion of strong reflection points, etc.). Simultaneously, historical trajectory features are extracted, using the centroid movement vector and velocity estimation from the past H frames (e.g., H=5 frames) as time-series input. If there are driving environment reference directions such as lane lines, road edges, or high-precision maps, these are included as additional input dimensions in the feature system. The pre-trained estimation network adopts a dual-branch fusion architecture. Branch A is the geometric and component feature branch, consisting of several fully connected layers, normalization layers, and activation layers (using ReLU or LeakyReLU activation functions) connected in series, followed by an attention weight module to adaptively weight the features of each component to highlight the contribution of effective features. Branch B is the velocity... The degree and environmental reference branches take historical speed trajectories and direction differences as input, and encode temporal correlations through GRU or LSTM networks. The output features of the two branches are concatenated and fed into a fusion fully connected layer. Finally, the output layer outputs the predicted heading angle of the corresponding target point cloud cluster (covering the range of -180° to +180° or 0° to 360°) and the uncertainty of the estimation accuracy (such as the standard deviation of the heading angle estimate). During the network training phase, a dataset of real lanes or test sites needs to be collected, covering stationary vehicles (different orientations 0° and 0°). Point cloud data with various angles (°, 45°, 90°, 135°, 180°, etc.), moving vehicles (speed 5-30m / s), different viewpoints, different reflection intensities, and occlusion scenarios were collected. The true heading angle θ_truth was obtained through manual annotation or visual / LiDAR-assisted annotation. The dataset was divided into training, validation, and test sets, and data augmentation operations were performed, including point cloud sparsification (simulating long-distance or low-reflection scenarios), adding noise, occlusion simulation, and simulation of tunnel or low-light conditions at night, to improve the network's generalization ability.

[0041] S105. For multiple valid heading angle assumptions for the same target point cloud cluster, select the heading angle prediction value with the smallest uncertainty as the final heading angle estimation result of the target point cloud cluster in the current frame.

[0042] For example, for the same target point cloud cluster, if multiple valid heading angle hypotheses (e.g., P=2) are retained after the aforementioned filtering, each valid hypothesis is input into the estimation network, which outputs a heading angle prediction value and its corresponding uncertainty (standard deviation). The heading angle prediction value with the smallest uncertainty is selected as the final heading angle estimation result for the target point cloud cluster in the current frame. If the uncertainties of multiple hypotheses are very close, they can be further combined with the cumulative scores of historical frames for weighted fusion.

[0043] After outputting the heading angle prediction and uncertainty for each frame, an error monitoring mechanism is activated. An uncertainty threshold is set (e.g., 10°). If the uncertainty exceeds this threshold for N_sigma frames (e.g., N_sigma=3), the current estimate is marked as low confidence. At the same time, a geometric direction deviation threshold is set. If the deviation between the heading angle prediction and the target geometric fitting direction exceeds this threshold for N_geo frames (e.g., N_geo=3), the geometric direction is determined to be inconsistent with the estimation result.

[0044] The self-calibration mechanism includes three core operations. The first is external parameter yaw correction, which addresses the fixed yaw error δ present in the radar installation. When continuous deviations exceed the threshold, the δ value is estimated using the mean or weighted average method, and this fixed error is deducted from all subsequent heading angle estimates. The second is feature and component weight adjustment. If a component has unstable micro-Doppler features or low point cloud density in multiple frames, the feature weight of that component is reduced; conversely, the weight of stable components is increased. The third is environmental reference auxiliary correction. When visual or map references are available and reliable (e.g., sufficient camera image brightness and small map matching error), they are used as additional constraints to increase the weight of that branch in network fusion, making the output result more consistent with the driving environment reference direction and achieving dynamic optimization of heading angle estimation.

[0045] The vehicle-mounted millimeter-wave radar target heading angle estimation method described in this technical solution obtains a reliable heading estimate even when the target is stationary or moving slowly by extracting geometric and micro-Doppler features. This overcomes the shortcomings of traditional methods that rely on speed or trajectory direction and perform poorly in low-speed scenarios. Simultaneously, by employing hypothesis generation, multi-criteria scoring, and dynamic elimination mechanisms, it possesses multi-candidate judgment capabilities without relying on a single hypothesis, thereby enhancing estimation stability in challenging environments such as sparse point clouds or occlusion. Furthermore, a temporal network with an attention mechanism is introduced to fuse, weight, and learn temporal correlations of geometric features, micro-Doppler features, reflection intensity features, and motion temporal information from continuous frame data, improving heading angle estimation accuracy and instantaneous response capability. In addition, this solution not only outputs the predicted heading angle value but also simultaneously outputs uncertainty or confidence indices representing the estimation accuracy, facilitating downstream modules to make decisions based on reliability. When the estimation confidence is low or the error persists, a self-correction mechanism is triggered, including extrinsic parameter yaw correction, component feature weight adjustment, and environmental reference auxiliary correction, thereby improving the stability and reliability of the system during long-term operation.

[0046] The target heading angle estimation device for vehicle-mounted radar provided by the present invention is described below. The target heading angle estimation device for vehicle-mounted radar described below can be referred to in correspondence with the target heading angle estimation method for vehicle-mounted radar described above.

[0047] Figure 2This is a schematic diagram of the target heading angle estimation device for vehicle-mounted radar provided by the present invention. For example, please refer to [link / reference]. Figure 2 As shown, the target heading angle estimation device of the vehicle-mounted radar may include: The data acquisition module is used to acquire raw point cloud data of continuous frames from the vehicle-mounted radar.

[0048] The target point cloud cluster acquisition module is used to perform clustering processing on the raw point cloud data to obtain at least one target point cloud cluster.

[0049] The effective heading angle hypothesis acquisition module is used to generate multiple candidate heading angle hypotheses for each target point cloud cluster in the initial frame. For each non-initial frame, based on the degree of fit between each candidate heading angle hypothesis and the geometry of the target point cloud cluster, the consistency with the direction of the centroid movement vector of the target point cloud cluster, and the consistency with the reference direction of the driving environment, the multiple candidate heading angle hypotheses of each target point cloud cluster are scored independently. Based on the scoring results, the multiple candidate heading angle hypotheses of each target point cloud cluster are independently filtered to obtain at least one effective heading angle hypothesis for each target point cloud cluster.

[0050] The prediction module is used to extract features from each target point cloud cluster to obtain the feature vector of each target point cloud cluster. For each effective heading angle hypothesis of each target point cloud cluster, the feature vector of the corresponding target point cloud cluster is input into the pre-trained estimation network, and the estimation network maps the feature vector to the heading angle prediction value and uncertainty of the corresponding effective heading angle hypothesis.

[0051] The target heading angle estimation module is used to select the heading angle prediction value with the smallest uncertainty from multiple valid heading angle assumptions for the same target point cloud cluster as the final heading angle estimation result of the target point cloud cluster in the current frame.

[0052] Specific limitations regarding the target heading angle estimation device for vehicle-mounted radar can be found in the above-mentioned limitations on target heading angle estimation for vehicle-mounted radar, and will not be repeated here. Each module in the aforementioned target heading angle estimation device for vehicle-mounted radar can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in hardware or independently of the processor in a computer device, or stored in software in the memory of a computer device, so that the processor can call and execute the corresponding operations of each module.

[0053] This specification also provides a computer-readable storage medium storing a computer program that can be used to execute the above-described... Figure 1 The provided method for estimating the target heading angle of vehicle-mounted radar.

[0054] This instruction manual also provides Figure 3 The schematic diagram of the computer device shown is as follows: Figure 3 As shown, at the hardware level, this computer device includes a processor, internal bus, network interface, memory, and non-volatile memory, and may also include other hardware required for business operations. The processor reads the corresponding computer program from the non-volatile memory into memory and then executes it to achieve the above. Figure 1 The provided method for estimating the target heading angle of vehicle-mounted radar.

[0055] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0056] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A method for estimating the target heading angle of a vehicle-mounted radar, characterized in that, include: Acquire raw point cloud data of consecutive frames from the vehicle-mounted radar; Clustering is performed on the raw point cloud data to obtain at least one target point cloud cluster; Within the initial frame, multiple candidate heading angle hypotheses are generated for each target point cloud cluster; For each non-initial frame, based on the degree of fit between each candidate heading angle hypothesis and the geometry of the target point cloud cluster, the consistency with the direction of the centroid movement vector of the target point cloud cluster, and the consistency with the reference direction of the driving environment, multiple candidate heading angle hypotheses for each target point cloud cluster are independently scored. Based on the scoring results, multiple candidate heading angle hypotheses for each target point cloud cluster are independently filtered to obtain at least one valid heading angle hypothesis for each target point cloud cluster. For each target point cloud cluster, feature extraction is performed to obtain the feature vector of each target point cloud cluster; for each effective heading angle hypothesis of each target point cloud cluster, the feature vector of the corresponding target point cloud cluster is input into the pre-trained estimation network, and the estimation network maps the feature vector to the heading angle prediction value and uncertainty of the corresponding effective heading angle hypothesis; For multiple valid heading angle assumptions for the same target point cloud cluster, the heading angle prediction with the smallest uncertainty is selected as the final heading angle estimation result of the target point cloud cluster in the current frame.

2. The target heading angle estimation method as described in claim 1, characterized in that, Before performing clustering processing on the raw point cloud data, preprocessing of the raw point cloud data is included, specifically including: Based on the preset radar detection distance threshold and the effective detection range of the road, the raw point cloud data that exceeds the distance constraint is filtered out; Filter out raw point cloud data with radar reflection intensity below the radar reflection intensity threshold; Filter out raw point cloud data whose velocity amplitude is lower than the velocity amplitude threshold and whose velocity variance is higher than the preset fluctuation threshold; Static clutter point clouds generated by ground reflectors or fixed structures can be filtered out by employing static background subtraction algorithms or ground plane filtering algorithms.

3. The target heading angle estimation method as described in claim 1, characterized in that, The assumption that multiple candidate heading angles for each target point cloud cluster are scored independently includes: For each target point cloud cluster, the principal axis direction of the target point cloud cluster is determined by principal component analysis, and the deviation between each candidate heading angle hypothesis of the target point cloud cluster and the principal axis direction is used as the geometric fit score of the candidate heading angle hypothesis. Obtain the centroid movement vector of each target point cloud cluster, and use the deviation between each candidate heading angle hypothesis and the centroid movement vector as the velocity direction consistency score of the candidate heading angle hypothesis; The deviation between each candidate heading angle assumption and the driving environment reference direction is used as the environmental reference consistency score of the candidate heading angle assumption; Based on the preset scoring weights, the geometric fitting score, velocity direction consistency score, and environmental reference consistency score are weighted and summed to obtain the frame score for each candidate heading angle hypothesis.

4. The target heading angle estimation method as described in claim 3, characterized in that, The process of independently filtering multiple candidate heading angle hypotheses for each target point cloud cluster based on the scoring results specifically includes: Set a preset number of sliding window frames, and for each candidate heading angle hypothesis of each target point cloud cluster, accumulate the frame scores of the candidate heading angle hypothesis in all frames within the sliding window to obtain the cumulative score of the candidate heading angle hypothesis; Determine whether the cumulative score of each candidate heading angle hypothesis is lower than a preset elimination threshold, or determine whether the candidate heading angle hypothesis is the lowest score within the window within a preset number of consecutive frames; If any of the above conditions are met, the candidate heading angle assumptions are eliminated; the remaining candidate heading angle assumptions in each target point cloud cluster are retained as the effective heading angle assumptions for the target point cloud cluster.

5. The target heading angle estimation method as described in claim 3, characterized in that, The step of extracting features from each target point cloud cluster to obtain the feature vector of each target point cloud cluster specifically includes: PCA analysis was performed on the preprocessed target point cloud cluster to obtain the principal component directions of the major and minor axes. At the same time, a rotation bounding rectangle was fitted to the target point cloud cluster and the side directions of the rotation bounding rectangle were obtained. Extract the geometric features of the target point cloud cluster, including the directional distribution of the normals of the edge points of the target point cloud cluster, aspect ratio, left-right symmetry and front-back symmetry of the graphic; Statistical histograms of radial velocity distribution are plotted for each component of the target point cloud cluster, and micro-Doppler features are extracted. The micro-Doppler features include the maximum energy range of the histogram, the velocity distribution width at the energy concentration point, the energy ratio of the positive and negative velocity ranges, the velocity distribution skewness, and the kurtosis. Extract the reflection intensity features of each component of the target point cloud cluster. The reflection intensity features include the average reflection intensity, the maximum reflection intensity, and the proportion of strong reflection points that are stronger than a preset threshold. Extract the positional features of strong reflection points, including the deviation of strong reflection points from the boundary of the rotating enclosing rectangle, the deviation of strong reflection points from the geometric center of the target cloud cluster, and the orientation concentration features of strong reflection points. After normalizing the geometric features, micro-Doppler features, reflection intensity features, and strong reflection point location features, the feature vectors of each target point cloud cluster are obtained by concatenating them.

6. The target heading angle estimation method as described in claim 3, characterized in that, The step of inputting the feature vector into a pre-trained estimation network, and having the estimation network map the feature vector to the predicted heading angle and uncertainty corresponding to the effective heading angle assumption, specifically includes: For each effective heading angle assumption of each target point cloud cluster, the feature vector of the target point cloud cluster in the current frame, the centroid movement vector of the past preset number of frames, and the driving environment reference direction are used as inputs to the pre-trained estimation network. The estimation network first extracts features and adds attention weights to the feature vectors through the geometric feature branch, and then performs temporal encoding on the centroid movement vector time series and the driving environment reference direction through the motion feature branch; The output features of the geometric feature branch and the motion feature branch are concatenated and fused. After mapping through a fully connected layer, the predicted heading angle corresponding to the effective heading angle assumption is output. At the same time, the uncertainty index, which characterizes the accuracy of the predicted value estimation, is output. The uncertainty index is the standard deviation of the heading angle estimate.

7. The target heading angle estimation method as described in claim 3, characterized in that, The pre-trained estimation network is a dual-branch fusion neural network that incorporates error monitoring and self-correction mechanisms. The dual-branch fusion neural network includes a geometric component feature branch and a velocity environment reference branch. The geometric component feature branch consists of a fully connected layer, a normalization layer, an activation layer, and an attention weight module, which is used to extract the feature vector of the target point cloud cluster by weight. The velocity environment reference branch consists of a GRU or LSTM network, which is used to encode the temporal feature of the centroid movement vector time series and the driving environment reference direction. The error monitoring mechanism monitors the heading angle estimation uncertainty and geometric direction deviation by setting a threshold. If the uncertainty exceeds the uncertainty threshold for multiple consecutive frames, or if the deviation between the heading angle prediction and the geometric fitting direction exceeds the geometric direction deviation threshold, the heading angle estimation is marked as a low confidence estimation. The self-correction mechanism includes extrinsic yaw correction, component feature weight adjustment, and environmental reference auxiliary correction. Extrinsic yaw correction is used to estimate and eliminate fixed yaw errors caused by radar installation. Component feature weight adjustment is used to reduce the feature weight of components with unstable features and low point cloud density. Environmental reference auxiliary correction is used to increase the weight of the driving environment reference direction in network fusion when the driving environment reference direction is reliable.

8. A target heading angle estimation device for vehicle-mounted radar, characterized in that, include: The data acquisition module is used to acquire raw point cloud data of continuous frames from the vehicle-mounted radar. The target point cloud cluster acquisition module is used to perform clustering processing on the raw point cloud data to obtain at least one target point cloud cluster. The effective heading angle hypothesis acquisition module is used to generate multiple candidate heading angle hypotheses for each target point cloud cluster within the initial frame; For each non-initial frame, based on the degree of fit between each candidate heading angle hypothesis and the geometry of the target point cloud cluster, the consistency with the direction of the centroid movement vector of the target point cloud cluster, and the consistency with the reference direction of the driving environment, multiple candidate heading angle hypotheses for each target point cloud cluster are independently scored. Based on the scoring results, multiple candidate heading angle hypotheses for each target point cloud cluster are independently filtered to obtain at least one valid heading angle hypothesis for each target point cloud cluster. The prediction module is used to extract features from each target point cloud cluster to obtain the feature vector of each target point cloud cluster. For each effective heading angle hypothesis of each target point cloud cluster, the feature vector of the corresponding target point cloud cluster is input into the pre-trained estimation network, and the estimation network maps the feature vector to the heading angle prediction value and uncertainty of the corresponding effective heading angle hypothesis. The target heading angle estimation module is used to select the heading angle prediction value with the smallest uncertainty from multiple valid heading angle assumptions for the same target point cloud cluster as the final heading angle estimation result of the target point cloud cluster in the current frame.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the target heading angle estimation method for vehicle-mounted radar as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the target heading angle estimation method for vehicle-mounted radar as described in any one of claims 1 to 7.

Images