A millimeter-wave radar traffic target clustering method based on prediction error feedback

Abstract

This invention discloses a millimeter-wave radar traffic target clustering method based on prediction error feedback. It employs weighted Euclidean distance, fully utilizing the spatial and Doppler information of the point cloud, to perform DBSCAN clustering as the initial clustering result. Based on relevant thresholds, it determines whether the single-target condition is met, fully utilizing temporal features. It performs Kalman filtering for five consecutive frames, assuming both multi-target and single-target scenarios, and uses prediction error feedback clustering as the final clustering result. This method leverages the small error between the predicted and actual states of large targets across consecutive frames to accurately cluster the point cloud, avoiding the phenomenon of single-frame clustering dividing the same target into multiple clusters, thus improving the accuracy of point cloud clustering for large targets.

Description

Technical Field

[0001] This invention belongs to the field of radar signal processing technology, specifically relating to a technology for target clustering and tracking of millimeter-wave radar point clouds based on prediction error feedback. Background Technology

[0002] In recent years, with the increasing frequency of traffic congestion and accidents, the market demand for intelligent transportation systems has been expanding. Target tracking is a crucial part of intelligent transportation systems. Before tracking, clustering techniques are needed to segment point cloud data into different clusters. Inadequacies in clustering algorithms can lead to the splitting or merging of tracking trajectories. Therefore, it is necessary to improve the clustering methods of millimeter-wave radar to enhance target tracking accuracy.

[0003] Existing research on point cloud clustering for millimeter-wave radar largely focuses on improving clustering algorithms for single-frame data. Furthermore, most clustering algorithms only utilize the spatial features of the point cloud, failing to fully leverage its Doppler characteristics. In traffic scenarios, a large vehicle target may reflect only its head, tail, or sides due to variations in its reflection points, and the point cloud distribution of the same target may differ between consecutive frames. Single-frame data clustering can lead to clustering interruptions, causing a single target to be clustered into multiple clusters, failing to correctly cluster sparsely distributed target point clouds, resulting in poor clustering results and impacting subsequent tracking. Summary of the Invention

[0004] The technical problem this invention aims to solve is that when millimeter-wave radar detects large targets, the resulting point clouds only contain different parts of the target. Traditional clustering algorithms tend to cluster a single large target into multiple targets, while this invention can correctly cluster point clouds from dispersed parts into a single target. To address this, this invention provides a millimeter-wave radar traffic target clustering method based on prediction error feedback. It employs weighted Euclidean distance, fully utilizing the spatial and Doppler information of the point cloud, and performs DBSCAN clustering as the initial clustering result. Based on relevant thresholds, it determines whether the single-target condition is met, fully utilizing temporal features. It performs Kalman filtering for five consecutive frames, assuming both multi-target and single-target scenarios, and uses prediction error feedback clustering as the final clustering result. This effectively improves the clustering accuracy of sparsely distributed point clouds.

[0005] The technical solution to achieve the objective of this invention is:

[0006] A millimeter-wave radar traffic target clustering method based on prediction error feedback includes the following steps:

[0007] Step 1: Input point cloud data and use the silhouette coefficient as the fitness function of the optimization algorithm. The formula for the silhouette coefficient s(i) is:

[0008]

[0009] In the formula, a(i) represents the intra-class distance, and b(i) represents the inter-class distance;

[0010] The optimal Eps and the optimal weight w for the weighted Euclidean distance are obtained through a grid search algorithm. The formula for the weighted Euclidean distance is:

[0011]

[0012] In the formula, x, y, and v represent the abscissa, axial ordinate, and velocity of the point cloud relative to the radar, respectively.

[0013] Step 2: Perform DBSCAN clustering of the point cloud using weighted Euclidean distance to obtain preliminary clustering results;

[0014] Step 3: Calculate the centroid and average velocity of each cluster after preliminary clustering. The calculation formula is:

[0015]

[0016] In the formula, n is the number of point clouds in the cluster;

[0017] Step 4: Determine the average velocity of each point cluster. If the difference is less than the threshold α, then check if the distance between the centroids of the point clusters is less than the threshold β. If it is, continue to step 5.

[0018] Otherwise, skip to step 10;

[0019] Step 5: Use a constant velocity (CV) model to describe the system's state vector. F(k) is the system state transition matrix, I(k) is the process noise distribution matrix, v(k) is the process noise vector, and the target state formula is:

[0020] X(k+1)=F(k)X(k)+I(k)v(k)

[0021] The point clusters that meet the conditions are treated as single-objective and multi-objective cases respectively, and state prediction is performed. G(k) is the input control term matrix, u(k) is the known input or control signal, and the prediction formula is:

[0022]

[0023] Step 6: Perform nearest neighbor association on the next frame and record the number of unassociated frames when the point cluster is regarded as a single target and the number of multiple targets respectively. If the number of unassociated frames is equal in both cases, continue to step 7.

[0024] If the number of unassociated frames in the multi-target case is greater than the number of unassociated frames in the single-target case, skip to step 9;

[0025] Otherwise, skip to step 10;

[0026] Step 7: Accumulate the prediction error value Er after 5 consecutive successful associations. The prediction error formula is:

[0027]

[0028] Step 8: Compare the prediction errors for the single-target case and the multi-target case. If the prediction error for the multi-target case is greater than the prediction error for the single-target case, continue to step 9.

[0029] Otherwise, skip to step 10;

[0030] Step 9: Merge the point clusters into one class as the final clustering result, and jump to step 11;

[0031] Step 10: Use the preliminary clustering results as the final clustering results;

[0032] Step 11: Clustering ends, output the results.

[0033] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0034] 1. First, a threshold is used to determine whether the target is suspected to be a large target. Kalman filtering is performed on both single-target and multi-target scenarios, and the clustering results are fed back using the prediction error.

[0035] 2. By taking advantage of the small error between the predicted state and the actual state of large targets in consecutive frames, the point cloud is accurately clustered, avoiding the phenomenon of single-frame clustering dividing the same target into multiple clusters, thus improving the accuracy of point cloud clustering for large targets. Attached Figure Description

[0036] Figure 1 This is the overall flowchart of the present invention;

[0037] Figure 2 For experimental scenarios;

[0038] Figure 3 This is the result of traditional single-frame DBSCAN clustering;

[0039] Figure 4 This is the clustering result of the method of the present invention. Detailed Implementation

[0040] The present invention will be further described below with reference to the accompanying drawings and embodiments, but this is not intended to limit the scope of the invention.

[0041] Example

[0042] A millimeter-wave radar traffic target clustering method based on prediction error feedback, referring to Figure 1 It includes the following steps:

[0043] Step 1: First, input the point cloud data and use the silhouette coefficient as the fitness function of the optimization algorithm. The silhouette coefficient s(i) is expressed by the following formula:

[0044]

[0045] Let x, y, and v be the abscissa, ythix, and velocity of the point cloud relative to the radar, respectively. Since the positional features of the point cloud are the primary clustering reference, the weights of the x and y features are assumed to be 1. The weighted Euclidean distance formula is expressed as:

[0046]

[0047] The optimal Eps was found to be 1.5, and the optimal weight w for the weighted Euclidean distance was 0.1, obtained through a grid search algorithm.

[0048] Step 2, in Figure 2 In the experimental scenario, there was only one bus target within the radar measurement range. Weighted Euclidean distance was used to perform DBSCAN clustering of the point cloud to obtain preliminary clustering results. The preliminary clustering would cluster the bus target into 2 point clusters.

[0049] Step 3: Calculate the centroid and average velocity of each cluster after preliminary clustering. Where n is the number of point clouds in the cluster, the calculation formula is:

[0050]

[0051] Step 4: Determine the average velocity of each point cluster. If the difference is less than the threshold α, then check if the distance between the centroids of the point clusters is less than the threshold β. If it is, continue to step 5; otherwise, jump to step 10. The values ​​of α and β are 1.3 and 7.5 respectively, based on experience with multiple large target point cloud clusters.

[0052] Step 5: Use a constant velocity (CV) model to describe the system's state vector. F(k) is the system state transition matrix, I(k) is the process noise distribution matrix, v(k) is the process noise vector, and the target state equation is:

[0053] X(k+1)=F(k)X(k)+I(k)v(k)

[0054] The point clusters that meet the conditions are treated as single-objective and multi-objective cases respectively, and state prediction is performed. G(k) is the input control term matrix, u(k) is the known input or control signal, and the prediction formula is:

[0055]

[0056] Step 6: Perform nearest neighbor association on the next frame and record the number of unassociated frames A in the cases of treating the point cluster as a single target and multiple targets respectively. The number of unassociated frames in the case of multiple targets is denoted as A_multi and the number of unassociated frames in the case of a single target is denoted as B_single. If the number of unassociated frames in the two cases is equal, that is, A_multi=B_single, then continue to step 7.

[0057] If the number of unassociated frames in the multi-target case is greater than the number of unassociated frames in the single-target case, i.e., A_multi > B_single, jump to step 9; otherwise, jump to step 10.

[0058] Step 7: Accumulate the prediction error value Er after 5 consecutive successful associations. The prediction error formula is:

[0059]

[0060] Step 8: Compare the prediction errors of the single-target case and the multi-target case. The prediction error value of the multi-target case is denoted as Er_multi, and the prediction error value of the single-target case is denoted as Er_single. If the prediction error of the multi-target case is greater than the prediction error of the single-target case, i.e. Er_multi>Er_single, continue to step 9.

[0061] Otherwise, skip to step 10.

[0062] Step 9: Merge the point clusters into one class as the final clustering result, and then proceed to step 11.

[0063] Step 10: Using the preliminary clustering result as the final clustering result, skip to step 11.

[0064] Step 11: Clustering ends, output the results.

[0065] The beneficial effects of this invention can be further illustrated through experimental simulation:

[0066] (1) Introduction to Experimental Scenarios and Equipment

[0067] Experimental scenarios such as Figure 2 As shown, point cloud data of traffic targets is collected using millimeter-wave radar. The radar is set up on the side of the road and mainly detects traffic targets within 80m. The targets detected by the millimeter-wave radar are compared with those detected by the camera, and the collected data is processed by algorithms.

[0068] (2) Simulation results

[0069] The clustering results of the traditional single-frame DBSCAN clustering algorithm are as follows: Figure 3 As shown, since the bus point cloud is roughly distributed at the head and tail, and the distance between them exceeds the neighborhood search radius, they are clustered into two clusters. Different shapes in the figure represent different target clusters. The final clustering result after combining prediction error feedback from multiple frames, assuming both single-target and multi-target scenarios, is shown below. Figure 4 As shown, bus targets can be correctly clustered, improving the clustering accuracy of large transportation targets.

Claims

1. A millimeter wave radar traffic target clustering method based on prediction error feedback, comprising the following steps: Step 1, input point cloud data, use the silhouette coefficient as the fitness function of the optimization algorithm, the formula of the silhouette coefficient s(i) is: ； In the formula, a(i) represents the intra-class distance, and b(i) represents the inter-class distance; Through the grid search algorithm, the optimal Eps and the optimal weight w of the weighted Euclidean distance are obtained, and the formula of the weighted Euclidean distance is: ； In the formula, x, y, and v are the horizontal coordinate, the axis coordinate, and the speed of the point cloud relative to the radar, respectively; Step 2, use the weighted Euclidean distance to perform point cloud DBSCAN clustering to obtain a preliminary clustering result; Step 3, calculate the Centroid of each cluster of points after initial clustering and the average velocity The formula is: ； In the formula, n is the number of point clouds in the cluster; Step 4, judge the average speed of each point cluster whether the phase difference is less than a threshold value a, if yes, continue to judge whether the distance of the point cluster center is less than a threshold value b, if yes, continue to step 5; otherwise, jump to step 10; Step 5, the state vector of the system is described using the uniform CV model , F(k) is the system state transition matrix, I(k) is the process noise distribution matrix, v(k) is the process noise vector, and the target state formula is: ； The point clusters meeting the conditions are regarded as single targets and multiple targets respectively, state prediction is performed, G(k) is the input control matrix, u(k) is the known input or control signal, and the prediction formula is: ； Step 6, perform nearest neighbor association on the next frame, and record the number of unassociated frames in the case of regarding the point cluster as a single target and a multiple target respectively, if the number of unassociated frames of the two is equal, step 7 is continued; If the number of unassociated frames in the case of multiple targets is greater than that in the case of single targets, step 9 is jumped to; otherwise, step 10 is jumped to; Step 7, accumulate the prediction error value after the correlation is successful for 5 consecutive frames The prediction error formula is: ； Step 8, compare the prediction errors in the case of single targets and multiple targets, if the prediction error in the case of multiple targets is greater than that in the case of single targets, step 9 is continued; otherwise, step 10 is jumped to; Step 9, combine the point clusters as a class, as the final clustering result, and jump to step 11; Step 10, take the preliminary clustering result as the final clustering result; Step 11, the clustering is ended, and the result is output.

Images