Point cloud clustering method based on depth map
The point cloud clustering method using depth maps addresses the inefficiencies of existing methods by employing linear compensation and adaptive clustering, enhancing speed and robustness for autonomous driving systems.
Patent Information
- Application Number
- CN202411270242.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-11
- Publication Date
- 2024-12-27
AI Technical Summary
Existing point cloud clustering methods for autonomous driving are time-consuming, not robust to noise, and fail to meet real-time requirements, especially in large-scale scenes, and are limited by the type of radar used or require fixed grid parameters.
A point cloud clustering method based on depth maps, involving linear fitting compensation, ground plane establishment, and adaptive clustering radius thresholds, which converts point clouds into depth maps for efficient clustering.
The method significantly improves clustering speed and real-time performance, adapting to varying point cloud densities and enhancing robustness to noise, while being applicable to both mechanical and solid-state laser radars.
Smart Images
Figure CN119206276A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of autonomous driving, and in particular to a point cloud clustering method based on a depth map. Background Art
[0002] Environmental perception is a key technology for autonomous driving. When an unmanned vehicle is driving, there are inevitably obstacles such as pedestrians, vehicles, and trees in the driving environment. The rapid and accurate detection of obstacles directly affects the driving safety of the unmanned vehicle.
[0003] At present, laser radar is generally used to scan the environment, collect point clouds, and cluster the point clouds to obtain the location of obstacles in the driving environment. In the prior art, point cloud clustering generally adopts the following three methods:
[0004] 1. Cluster the point cloud according to its features, such as distance, density, etc., for example, using Euclidean clustering, KMEANS, DBSCAN, etc. However, this method classifies the point clouds in the neighborhood with high similarity in distance or other features to the point cloud into the same category. There are a lot of neighborhood searches in the method, but due to the disorder of the point cloud, the search is time-consuming. Even if the kdtree method is used, it is difficult to meet the real-time requirements of large-scale point cloud scenes, and it is not robust to noise in the original point cloud.
[0005] 2. Cluster the laser beam Scan-Line Run (SLR), and cluster each scan line in turn according to the rotation characteristics of the mechanical radar. Although this method uses traversal processing of the laser radar beam and directly compares the features of adjacent points without the need for neighborhood search, which improves real-time performance, this method has major limitations. It is too dependent on the scanning characteristics of the laser radar and is only applicable to mechanical laser radars, not solid-state laser radars.
[0006] 3. Cluster the grids, project the point cloud to horizontal rectangular coordinates or polar coordinate grids, and then cluster the grids. This method also does not require neighborhood search, and can directly find nearby grids based on indexes, with high real-time performance. However, the grid unit parameters in this method cannot be adaptively adjusted, which can easily cause under-segmentation or over-segmentation. Summary of the invention
[0007] In view of the above-mentioned problems, the present invention provides a point cloud clustering method based on depth map, which clusters point clouds by depth map, greatly improves the clustering speed and ensures the real-time performance of the unmanned vehicle system.
[0008] The technical solution of the present invention is: a point cloud clustering method based on a depth map, comprising the following steps:
[0009] 1) The laser radar on the unmanned vehicle emits a laser beam to collect point cloud data, and the collected point cloud data is compensated by the linear fitting compensation method;
[0010] 2) Establish a ground plane, traverse the point cloud, and remove the point cloud data whose point cloud height is lower than the ground plane height in the compensated point cloud data;
[0011] 3) Establishing a depth map based on the remaining point cloud data above the ground plane in step 2), wherein each pixel point of the depth map includes distance data from the obstacle to the laser radar;
[0012] 4) Cluster each pixel according to the depth map, and perform neighborhood search for each pixel using the clustering radius threshold. If the distance between the pixel and its neighboring pixels is less than the clustering radius threshold, they are classified into one category, and all pixels are traversed to complete clustering;
[0013] 5) Back-project the clustered depth map to the point cloud to form a visual image.
[0014] Preferably, in step 1), the steps of the linear fitting compensation method are:
[0015] 1.1) Traverse the point cloud data collected by each laser in the laser radar;
[0016] 1.2) If there is missing point cloud data between two point cloud data collected by the same laser, and the distance between the two point cloud data is less than a preset threshold, a linear line is fitted according to the two point cloud data, and the missing point cloud data is supplemented by linear interpolation.
[0017] Furthermore, the preset threshold is 0.3m.
[0018] Preferably, in step 2), the step of establishing a ground plane is:
[0019] 2.1) Select n point cloud data with the lowest height from the point cloud data collected by the laser radar, and establish a ground plane equation, which is:
[0020] ax+by+cz+d=0
[0021] In the formula, a, b, c, and d are equation coefficients;
[0022] 2.2) Calculate the mean coordinates of the n point cloud data coordinates, and calculate the covariance matrix based on the mean coordinates;
[0023] 2.3) According to the covariance matrix of the n point cloud data, three singular vectors of the covariance matrix and the singular values corresponding to the three singular vectors are obtained through singular value decomposition SVD;
[0024] 2.4) Compare the sizes of the three singular values, select the smallest singular value, and use the singular vector corresponding to the smallest singular value as the normal vector of the ground plane equation;
[0025] 2.5) Determine the three coefficients a, b, and c of the ground plane equation based on the normal vector, then substitute the mean coordinate into the ground plane equation to determine the parameter d, and complete the establishment of the ground plane.
[0026] Furthermore, the covariance matrix can be calculated as follows:
[0027]
[0028] In the formula, C is the covariance matrix, S is the total number of point clouds, and s i is the coordinate of the i-th point cloud, are the mean coordinates.
[0029] Preferably, in step 4), the cluster radius threshold includes a horizontal cluster radius threshold and a vertical cluster radius threshold, and the calculation formulas of the cluster radius thresholds are:
[0030]
[0031] In the formula, range ori is the depth of the center pixel, range neighbor is the depth of the neighborhood pixel, θ is the angular resolution, the angular resolution θ of the clustering radius threshold in the horizontal direction is 0.133°, and the angular resolution θ of the clustering radius threshold in the vertical direction is 1.06°.
[0032] The advantages of the present invention are as follows: the present invention converts the point cloud into a depth map and then performs clustering processing, thereby ordering the disordered point cloud, simplifying the neighborhood search, and greatly improving the clustering speed. In addition, the clustering threshold and clustering radius of the present invention change with the distance of the laser radar, solving the problem that when the laser radar scans the point cloud data, the closer the point cloud data is to the laser radar, the higher the density, and the farther the point cloud data is from the laser radar, the lower the density, so that the present invention can adapt to the clustering requirements of different distances, enhancing the practicality of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS
[0033] Figure 1 is a flow chart of the present invention;
[0034] Figure 2 This is the point cloud image after compensation of the present invention;
[0035] Figure 3 This is a schematic diagram of the present invention after removing the point cloud below the ground plane;
[0036] Figure 4 This is the final visualized image of the present invention. DETAILED DESCRIPTION
[0037] See also Figure 1 , a point cloud clustering method based on depth map, comprising the following steps:
[0038] 1) The laser radar on the unmanned vehicle emits a laser beam to collect point cloud data, and the collected point cloud data is compensated by the linear fitting compensation method. The compensated results are as follows: Figure 2 ; The steps of the linear fitting compensation method are:
[0039] 1.1) Traverse the point cloud data collected by each laser in the laser radar;
[0040] 1.2) If there is missing point cloud data between two point cloud data collected by the same laser, and the distance between the two point cloud data is less than 0.3m, a linear line is fitted according to the two point cloud data, and the missing point cloud data is supplemented by linear interpolation.
[0041] 2) Establish a ground plane, traverse the point cloud, and remove the point cloud data whose point cloud height is lower than the ground plane height in the compensated point cloud data. The result after removing the point cloud data whose height is lower than the ground plane height is as follows: Figure 3 The steps to establish a ground plane are:
[0042] 2.1) Select the 100 point cloud data with the lowest height from the point cloud data collected by the laser radar, and establish a ground plane equation, which is:
[0043] ax+by+cz+d=0
[0044] In the formula, a, b, c, and d are equation coefficients;
[0045] 2.2) Calculate the mean coordinates of the 100 point cloud data coordinates, and calculate the covariance matrix based on the mean coordinates; the covariance matrix can be calculated as follows:
[0046]
[0047] In the formula, C is the covariance matrix, S is the total number of point clouds, and s i is the coordinate of the i-th point cloud, are the mean coordinates.
[0048] 2.3) According to the covariance matrix of the 100 point cloud data, three singular vectors of the covariance matrix and the singular values corresponding to the three singular vectors are obtained through singular value decomposition (SVD);
[0049] 2.4) Compare the sizes of the three singular values, select the smallest singular value, and use the singular vector corresponding to the smallest singular value as the normal vector of the ground plane equation;
[0050] 2.5) Determine the three coefficients a, b, and c of the ground plane equation based on the normal vector, then substitute the mean coordinate into the ground plane equation to determine the parameter d, and complete the establishment of the ground plane.
[0051] 3) Establishing a depth map based on the remaining point cloud data above the ground plane in step 2), wherein each pixel point of the depth map includes distance data from the obstacle to the laser radar;
[0052] 4) Cluster each pixel according to the depth map, and perform neighborhood search on each pixel through the clustering radius threshold. If the distance between the pixel and its neighboring pixel is less than the clustering radius threshold, they are divided into one category, and all pixels are traversed to complete clustering; the clustering radius threshold includes the horizontal clustering radius threshold and the vertical clustering radius threshold, and the calculation formula of the clustering radius threshold is:
[0053]
[0054] In the formula, range ori is the depth of the center pixel, range neighbor is the depth of the neighborhood pixel, θ is the angular resolution, the angular resolution θ of the clustering radius threshold in the horizontal direction is 0.133°, and the angular resolution θ of the clustering radius threshold in the vertical direction is 1.06°.
[0055] 5) See Figure 4 , the depth map after clustering is back-projected to the point cloud to form a visual image.
Claims
1. A point cloud clustering method based on depth map, characterized in that: The following steps are involved: 1) The laser radar on the unmanned vehicle emits a laser beam to collect point cloud data, and the collected point cloud data is compensated by the linear fitting compensation method; 2) Establish a ground plane, traverse the point cloud, and remove the point cloud data whose point cloud height is lower than the ground plane height in the compensated point cloud data; 3) Establishing a depth map based on the remaining point cloud data above the ground plane in step 2), wherein each pixel point of the depth map includes distance data from the obstacle to the laser radar; 4) Cluster each pixel according to the depth map, and perform neighborhood search on each pixel using the clustering radius threshold. If the distance between the pixel and its neighboring pixels is less than the clustering radius threshold, they are classified into one category, and all pixels are traversed to complete clustering; 5) Back-project the clustered depth map to the point cloud to form a visual image.
2. The point cloud clustering method based on depth map according to claim 1, characterized in that: In step 1), the steps of the linear fitting compensation method are: 1.1) Traverse the point cloud data collected by each laser in the laser radar; 1.2) If there is missing point cloud data between two point cloud data collected by the same laser, and the distance between the two point cloud data is less than a preset threshold, a linear line is fitted according to the two point cloud data, and the missing point cloud data is supplemented by linear interpolation.
3. The point cloud clustering method based on depth map according to claim 2, characterized in that: The preset threshold is 0.3m.
4. The point cloud clustering method based on depth map according to claim 1, characterized in that: In step 2), the steps to establish a ground plane are: 2.1) Select n point cloud data with the lowest height from the point cloud data collected by the laser radar, and establish a ground plane equation, which is: ax+by+cz+d=0 In the formula, a, b, c, and d are equation coefficients; 2.2) Calculate the mean coordinates of the n point cloud data coordinates, and calculate the covariance matrix based on the mean coordinates; 2.3) According to the covariance matrix of the n point cloud data, three singular vectors of the covariance matrix and the singular values corresponding to the three singular vectors are obtained through singular value decomposition SVD; 2.4) Compare the sizes of the three singular values, select the smallest singular value, and use the singular vector corresponding to the smallest singular value as the normal vector of the ground plane equation; 2.5) Determine the three coefficients a, b, and c of the ground plane equation based on the normal vector, then substitute the mean coordinate into the ground plane equation to determine the parameter d, and complete the establishment of the ground plane.
5. The point cloud clustering method based on depth map according to claim 4, characterized in that: The covariance matrix can be calculated as: In the formula, C is the covariance matrix, S is the total number of point clouds, and s i is the coordinate of the i-th point cloud, are the mean coordinates.
6. The point cloud clustering method based on depth map according to claim 1, characterized in that: In step 4), the cluster radius threshold includes a horizontal cluster radius threshold and a vertical cluster radius threshold, and the calculation formulas of the cluster radius thresholds are: In the formula, range ori is the depth of the center pixel, range neighbor is the depth of the neighborhood pixel, θ is the angular resolution, the angular resolution θ of the clustering radius threshold in the horizontal direction is 0.133°, and the angular resolution θ of the clustering radius threshold in the vertical direction is 1.06°.