Underground parking lot ground contour extraction method based on laser radar three-dimensional point cloud
By combining lidar inertial SLAM technology and image processing methods, the problems of noise interference and occlusion in wall line extraction in underground parking environments were solved, achieving high-precision wall and column extraction and improving the accuracy and efficiency of data acquisition.
Patent Information
- Application Number
- CN202410927965.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-11
- Publication Date
- 2024-11-22
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Existing technologies struggle to quickly and accurately extract wall lines in underground parking environments. In particular, noise interference and occlusion issues in complex environments lead to the loss of point cloud information for columns and walls. Furthermore, traditional methods are sensitive to noise or rely on prior conditions, making it difficult to identify short straight lines.
Using tightly coupled lidar inertial SLAM technology, combined with PCA, RANSAC and clustering algorithms, the ground point cloud is segmented, denoising is performed using octrees, and point clouds of walls and columns are extracted based on point cloud histograms and image processing methods. Gaussian filtering is used to smooth the wall curves to construct an accurate ground contour of the underground parking lot.
It achieves high-precision extraction of wall columns in underground parking lots, improves the accuracy and efficiency of data collection, reduces noise interference, and ensures the accuracy of wall line extraction in complex environments.
Smart Images

Figure CN119006645A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a method for accurately extracting a plane structure in an underground parking environment, and specifically to a method for extracting a ground contour of an underground parking lot based on a laser radar 3D point cloud. Background Art
[0002] As the number of cars continues to increase, the demand for parking has surged. The increasingly intensified contradiction between the supply and demand of parking spaces across the country has led to the emergence of smart parking. As the basis of smart parking services, how to build an accurate parking map has become a research hotspot for smart parking companies.
[0003] In parking lot environments (especially underground parking lots), the ground contours are mainly described by the wall lines of a large number of walls and columns. The rapid and accurate extraction of wall lines is of great significance for the real-time generation of parking lot maps. With the rapid development of mobile laser scanning and laser real-time positioning and mapping (SLAM) technology, backpack and handheld indoor 3D laser scanning based on LiDAR (Light Detection and Ranging, LiDAR) provide high-density and high-precision 3D laser point clouds for parking lot wall line extraction. Existing wall line extraction methods mostly use principal component analysis (PCA) to calculate point clouds. [1] Or is it Random Sample Consensus (RANSAC) [2] To extract the plane, the PCA algorithm is sensitive to noise points, and the RANSAC algorithm relies on the prior condition of the number of fitted lines. [3] As a classic straight line detection method, the extracted straight line cannot determine the endpoints and it is difficult to identify shorter wall lines.
[0004] In addition, for complex environments such as parking lots (especially underground parking lots), there are a lot of foreign object (vehicles, ceiling and ground point clouds, etc.) noise and reflection noise, which makes it difficult to intelligently segment the wall column point cloud; on the other hand, in the parking lot environment, due to the mutual occlusion between obstacles, some columns and walls may only be scanned to part of the surface when using LiDAR scanning. This occlusion problem leads to the loss of column and wall point cloud information, thereby increasing the difficulty of wall line extraction. In addition, if the route is not reasonably planned during the LiDAR point cloud data collection process, the edge and corner areas of the parking lot may not be fully covered, resulting in missing wall point cloud data in these areas. However, in order to improve the accuracy of data collection, if the edge area of the parking lot is to be fully scanned and detailed scans are performed around each column, this will significantly increase the time required for data collection.
[0005] [1]Anagnostopoulos I,Viorica Patraucean,Brilakis I et al.Detection of walls,floors,and ceilings in point cloud data[J].Construction ResearchCongress,2016:2302~2311.
[0006] [2]Schnabel R,Wahl R,Klein R.Efficient RANSAC for point cloud shapedetection[J].Computer Graphics
[0007] Forum, 2010, 26(2): 214~226.
[0008] [3] Sun Xianlong, Han Litao, Zhou Lijuan, et al. A building wall line extraction based on indoor point cloud data[J]. Surveying and Mapping Engineering, 2021, 30(2): 65-71. Summary of the invention
[0009] The present invention constructs a global three-dimensional point cloud map of an underground parking lot based on a tightly coupled laser radar inertial SLAM technology. The ground points are segmented, the ground plane is estimated, and the plumb direction of the point cloud map is parallel to the plumb direction of the world coordinate system. Relying on the octree data structure, fast denoising of the global point cloud is achieved. With PCA, the wall column point cloud is roughly segmented; based on the point cloud histogram, the point cloud slices are extracted. The ground points and ceiling points are removed; based on the clustering algorithm, the wall column point clusters are further clustered, the vehicle point cloud is removed, and projected to the XOY plane. Based on the discreteness of each point cluster, the column point cluster and the wall point cluster are preliminarily distinguished, and the column point cluster is secondary screened by clustering and PCA. For the column point cluster, the column wall line is fitted based on the RANSAC algorithm. For the wall point cluster, a point cloud binary image is constructed, and the expansion, corrosion, and refinement operations in image processing are used in turn to fit the wall line, and finally the wall curve is smoothed by Gaussian filtering.
[0010] In order to achieve the purpose of the present invention, the following technical solution will be adopted: a method for extracting the ground contour of an underground parking lot based on a laser radar 3D point cloud. It includes the following steps:
[0011] S1. Data collection. Calibrate and synchronize the LiDAR and the Inertial Measurement Unit (IMU), and collect the LiDAR 3D point cloud dataset and IMU measurement dataset of the underground parking environment.
[0012] S2. Build a 3D point cloud map. Based on the real-time positioning and mapping technology of LiDAR-IMU fusion, each frame of the collected 3D point cloud Perform registration to solve the 6D pose T of the acquisition device at the time of each LiDAR frame point cloud acquisition k ; Filter the LiDAR point cloud keyframes, and based on the solved 6D pose, stitch the point clouds of each keyframe to build a global 3D point cloud map of the underground parking lot
[0013] S3. Point cloud map plumb direction correction:
[0014] S3-1, segment the ground point cloud. Based on the LiDAR point cloud keyframes obtained in S2, the ground point cloud in each LiDAR keyframe is segmented using the ground segmentation algorithm. Based on the 6D pose of the current key frame solved in S2, the ground point clouds in each key frame are spliced to construct a global ground point cloud.
[0015] S3-2, iteratively estimate the ground plane equation. In the i-th iteration, randomly select N points in the global ground point cloud, assume that these N points are on the ground plane to be estimated, and use principal component analysis (PCA) to calculate the eigenvector (normal vector) of the plane. Traverse the global ground point cloud and calculate the distance between each ground point and the current estimated ground plane, which is less than the set threshold d thr The points are recorded as interior points, and the rest of the points are set as exterior points. The number of interior points is counted. Number of external points If the number of inliers estimated this time is greater than the maximum number of inliers in history Update the historical best ground plane model for When the number of iterations reaches the set upper limit, or the ratio of the current number of internal points to the total number of points is greater than a certain threshold, the iterative estimation ends and normalization is performed. Get the final ground plane feature vector
[0016] S3-3. Calculate the ground plane feature vector The angle between the unit z-axis vector and the global three-dimensional point cloud is obtained by solving the corresponding rotation matrix R. Perform corresponding rotation to make the plumb direction of the point cloud map parallel to the plumb direction of the world coordinate system, and obtain the corrected global point cloud map
[0017] S4. Noise preprocessing:
[0018] S4-1. Remove noise points based on distance. Store the global point cloud map in the form of an octree Traverse the global point cloud. For each sampling point p j , statistics are based on p j If the number of all points within a specific range is less than the set threshold, the current point is regarded as an outlier (noise point) and is filtered out.
[0019] S4-2. Rough segmentation of wall and column point cloud based on PCA. Restore the global point cloud map in the form of KDTree Traverse the global point cloud. For each sampling point p j , search for j The nearest K points are selected and PCA is used to obtain the normal vector of each sampling point. If the product of the normal vector and the unit z-axis is less than a certain threshold (that is, the normal vector and the z-axis are nearly perpendicular), the sampling point is considered to be a wall point and is retained. The remaining points are considered to be non-wall points and are filtered out as noise points.
[0020] S5. Clustering and segmentation of wall column point cloud:
[0021] S5-1. Point cloud slice extraction based on point cloud histogram. Construct a point cloud density histogram in the z-axis direction, observe its distribution, and intercept point cloud data of a certain thickness to remove ground points and ceiling points.
[0022] S5-2. Use the DBSCAN clustering algorithm to perform preliminary clustering of the wall and column point clouds in the parking lot. In order to achieve preliminary clustering of the wall and column point clouds while reducing the phenomenon of over-segmentation of point clusters and reducing the phenomenon of sparse point clusters being misjudged as noise points, the parameter settings are relatively conservative. Calculate the distribution range of the z axis in each point cluster. If the z value of the midpoint of the current point cluster is less than the set height threshold z thr The ratio is greater than the set ratio threshold ratio thr , then the point cluster is judged as a vehicle point cluster and removed. Finally, each point cluster is projected onto the XOY plane to obtain a two-dimensional wall column point cloud.
[0023] S6 wall column point cluster classification:
[0024] S6-1. Preliminary classification of wall column point clusters based on point cloud discreteness. Traverse each point cluster, and for the current point cluster, calculate the center point of the point cluster; traverse all points in the point cluster, calculate the distance between each point and the center point of the point cluster and add them up. If the sum result is less than the set threshold, it is regarded as the column point cluster C. col , otherwise, it is regarded as the wall point cluster C wall .
[0025] S6-2. Column point cluster C for preliminary classification col In order to solve the problem that some columns are too close to each other, resulting in incomplete cluster classification, the DBSCAN clustering algorithm is used to perform secondary clustering segmentation on each initially classified column point cluster to obtain the column point cluster In this process, the parameter setting is relatively radical. In order to avoid excessive segmentation of the column point clusters caused by radical parameter settings, the following segmentation rules are set: for each column, if the number of points in the newly generated point cluster after the secondary clustering segmentation is greater than the preset threshold and the number of point clusters that meet the conditions is greater than 2, then the secondary clustering segmentation is considered valid, otherwise the original point cluster is retained.
[0026] S6-3. Since the short straight-line wall point clusters are difficult to distinguish from the column point clusters due to their small discreteness, the PCA algorithm is used to segment the column point clusters after secondary clustering. Perform point distribution analysis to obtain the final column point cluster If the point cluster is linearly distributed, reclassify it as a short-line wall point cluster
[0027] S7, column curve extraction. Traverse each column point cluster Due to the mutual occlusion between obstacles, many column point clusters can only be scanned to part of the surface. To solve this problem, based on the random sampling consensus method (RANSAC), two mutually perpendicular straight line segments are fitted, and the four sides of the square are further constructed to solve its four vertices. The shape of the current column and the solved coordinate information are stored in the column information structure middle.
[0028] S8. Wall curve extraction:
[0029] S8-1. For each short-line wall point cluster Since only one straight line is fitted, the straight line is fitted using the RANSAC algorithm, and the endpoints of the straight line segment are determined according to the distribution range of the points in the point cluster and stored in the wall information structure.
[0030]
[0031] S8-2, traverse each wall point cluster C wall (non-linear point clusters), establish a two-dimensional grid, consider the wall width, set the grid spacing, and rasterize the point cloud data into a binary image based on whether there are points in the grid.
[0032] S8-3, dilation and erosion operations are performed on the binary image. The dilation operation is to convolve the binary image with the dilation operator, find the maximum value of the pixels in the area covered by the dilation operator, and assign the maximum value to the neighboring pixels, thereby increasing the highlight area. The erosion operation uses the erosion operator convolution, and its value is 1 only when the corresponding element values of the convolution kernel are all 1, otherwise it is 0.
[0033] S8-4. Use the Zhang-Suen thinning algorithm to iteratively extract the skeleton of the binary image, replace the pixel with the coordinates of the pixel center, and extract the coordinates of all highlighted pixels.
[0034] S8-5, connect the refined highlighted pixel coordinates and smooth the curve using Gaussian filtering. Finally, divide the smoothed wall line into point clusters and store them in the wall information structure. middle.
[0035] S9, curve drawing. Draw all the column information structures With wall information structure Draw the wall and column curves as individual solid elements. BRIEF DESCRIPTION OF THE DRAWINGS
[0036] Figure 1 Flowchart
[0037] Figure 2 Global view of 3D point cloud
[0038] Figure 3-1 / Figure 3-2 Z-axis histogram before and after ground segmentation correction
[0039] Figure 4 Cluster classification diagram
[0040] Figure 5 Column point cluster segmentation diagram
[0041] Figure 6 Wall point cluster segmentation map
[0042] Figure 7 Non-linear wall point cluster processing results
[0043] Figure 8 Column curve fitting and wall curve fitting diagram
[0044] Fig. 9 Comparison diagram of fitting results
[0045] Fig.10 CAD original drawing
[0046] Fig.11 Planar structural drawing of wall columns extracted from CAD
[0047] Fig.12 Wall line extraction accuracy evaluation experiment diagram DETAILED DESCRIPTION
[0048] The following describes the implementation of the present invention in detail with reference to the accompanying drawings and embodiments:
[0049] 1. Collection of experimental data
[0050] The experimental environment was set up in the underground parking lot of Yangtze River Delta Research Institute (Quzhou) of University of Electronic Science and Technology of China in Kecheng District, Quzhou City, Zhejiang Province. The pose calibration of LiDAR and IMU was completed, and three-dimensional LiDAR point cloud data and IMU measurement data were collected.
[0051] 2. Build a 3D point cloud map. Based on the real-time positioning and mapping technology of LiDAR-IMU fusion, each frame of the collected 3D point cloud Perform registration to solve the 6D pose T of the acquisition device at the time of each LiDAR frame point cloud acquisition k ; Filter the LiDAR point cloud keyframes, and based on the solved 6D pose, stitch the point clouds of each keyframe to build a global 3D point cloud map of the underground parking lot
[0052] 3. Point cloud map plumb direction correction
[0053] Since the coordinate system of the global 3D point cloud map constructed by the current LiDAR-IMU fusion SLAM technology is usually consistent with the coordinate system of the first LiDAR or IMU measurement frame, rather than the absolute world coordinate system, there is often a certain angle between the plumb bob direction of the constructed global 3D point cloud map and the unit z-axis direction in the world coordinate system. The plumb bob direction of the point cloud map needs to be corrected to be parallel to the unit z-axis direction of the world coordinate system, so that the 3D point cloud map can be conveniently projected onto the XOY plane.
[0054] (1) First, segment the ground point cloud. Based on the LiDAR point cloud keyframes obtained in S2, the ground point cloud in each LiDAR keyframe is segmented using a ground segmentation algorithm. Based on the 6D pose of the current key frame solved in step 2, the ground point clouds in each key frame are spliced to construct a global ground point cloud.
[0055] (2) Iteratively estimate the ground plane equation. In the i-th iteration, randomly select N points in the global ground point cloud. Assuming that these N points are on the ground plane to be estimated, use the principal component analysis (PCA) to calculate the eigenvector (normal vector) of the plane. Traverse the global ground point cloud and calculate the distance between each ground point and the current estimated ground plane, which is less than the set threshold d thr The points are recorded as interior points, and the rest of the points are set as exterior points. The number of interior points is counted. Number of external points If the number of inliers estimated this time is greater than the maximum number of inliers in history Update the historical best ground plane model for When the number of iterations reaches the set upper limit, or the ratio of the current number of internal points to the total number of points is greater than a certain threshold, the iterative estimation ends and normalization is performed. Get the final ground plane feature vector
[0056] (3) Calculate the ground plane feature vector The angle between the unit z-axis vector and the global three-dimensional point cloud is obtained by solving the corresponding rotation matrix R. Perform corresponding rotation to make the plumb direction of the point cloud map parallel to the plumb direction of the world coordinate system, and obtain the corrected global point cloud map
[0057] 4. Noise preprocessing
[0058] (1) Remove noise points based on distance. Store the global point cloud map in the form of an octree Traverse the global point cloud. For each sampling point p j , statistics are based on pj If the number of all points within a specific range is less than the set threshold, the current point is regarded as an outlier (noise point) and is filtered out.
[0059] (2) Rough segmentation of wall and column point cloud based on PCA. Restore the global point cloud map in the form of KDTree Traverse the global point cloud. For each sampling point p j , search for j The nearest K points are selected and PCA is used to obtain the normal vector of each sampling point. If the product of the normal vector and the unit z-axis is less than a certain threshold (that is, the normal vector and the z-axis are nearly perpendicular), the sampling point is considered to be a wall point and is retained. The remaining points are considered to be non-wall points and are filtered out as noise points.
[0060] 5. Clustering and segmentation of wall and column point clouds
[0061] (1) Point cloud slice extraction based on point cloud histogram: Construct a point cloud density histogram in the z-axis direction, observe its distribution, and cut off point cloud data of a certain thickness to remove ground points and ceiling points.
[0062] (2) Using the DBSCAN clustering algorithm, the wall and column point clouds in the parking lot are initially divided into point clusters; the z-axis distribution range of each point cluster is calculated; if the z value of the midpoint of the current point cluster is less than the set height threshold z thr The ratio is greater than the set ratio threshold ratio thr , then the point cluster is judged as a vehicle point cluster and removed. Finally, each point cluster is projected onto the XOY plane to obtain a two-dimensional wall column point cloud.
[0063] 6. Classification of wall column point clusters
[0064] (1) Preliminary classification of wall column point clusters based on point cloud discreteness. Traverse each point cluster, and for the current point cluster, calculate the center point of the point cluster; traverse all points in the point cluster, calculate the distance between each point and the center point of the point cluster and accumulate the sum. If the sum result is less than the set threshold, it is regarded as the column point cluster C. col , otherwise, it is regarded as the wall point cluster C wall .
[0065] (2) Targeting the initially classified column point cluster C col In order to solve the problem that some columns are too close to each other, resulting in incomplete cluster classification, the DBSCAN clustering algorithm is used to perform secondary clustering segmentation on each initially classified column point cluster to obtain the column point cluster In this process, the parameter setting is relatively radical. In order to avoid excessive segmentation of the column point clusters caused by radical parameter settings, the following segmentation rules are set: for each column, if the number of points in the newly generated point cluster after the secondary clustering segmentation is greater than the preset threshold and the number of point clusters that meet the conditions is greater than 2, then the secondary clustering segmentation is considered valid, otherwise the original point cluster is retained.
[0066] (3) Since the short straight-line wall point clusters are difficult to distinguish from the column point clusters due to their small discreteness, the PCA algorithm is used to segment the column point clusters after secondary clustering. Perform point distribution analysis to obtain the final column point cluster If the point cluster is linearly distributed, reclassify it as a linear wall point cluster
[0067] 7. Extract column curves. Traverse each column point cluster Due to the mutual occlusion between obstacles, many column point clusters can only be scanned to part of the surface. To solve this problem, based on the random sampling consensus method (RANSAC), two mutually perpendicular straight line segments are fitted, and the four sides of the square are further constructed to solve its four vertices. The shape of the current column and the solved coordinate information are stored in the column information structure middle.
[0068] 8. Wall curve extraction
[0069] (1) For each linear wall point cluster Since only one straight line is fitted, the straight line is fitted using the RANSAC algorithm, and the endpoints of the straight line segment are determined according to the distribution range of the points in the point cluster and stored in the wall information structure.
[0070] (2) Traverse each wall point cluster C wall (non-linear point clusters), establish a two-dimensional grid, consider the wall width, set the grid spacing, and rasterize the point cloud data into a binary image based on whether there are points in the grid.
[0071] (3) Perform dilation and erosion operations on the binary image. The dilation operation is to convolve the binary image with the dilation operator, find the maximum value of the pixels in the area covered by the dilation operator, and assign the maximum value to the neighboring pixels, thereby increasing the highlight area. The erosion operation uses the erosion operator convolution. When the values of the elements corresponding to the convolution kernel are all 1, its value is 1, otherwise it is 0.
[0072] (4) The skeleton of the binary image is iteratively extracted using the Zhang-Suen thinning algorithm, the coordinates of all highlighted pixels are extracted by replacing the pixel with the coordinates of the pixel center.
[0073] (5) Connect the refined highlighted pixel coordinates and smooth the curve using Gaussian filtering. Finally, the smoothed wall line is divided into point clusters and stored in the wall information structure. middle.
[0074] 9. Curve drawing. Draw all the column information structures With wall information structure Draw the wall and column curves as individual solid elements.
[0075] Example
[0076] This experiment was completed in the B1 underground parking lot of the Yangtze River Delta Research Institute (Quzhou) of University of Electronic Science and Technology of China in Kecheng District, Quzhou City, Zhejiang Province. The building area of this floor is 11,168 square meters, with 189 motor vehicle parking spaces, including 7 barrier-free parking spaces and 36 electric vehicle parking spaces.
[0077] During the experiment, RS-LiDAR-16 LiDAR from Suteng and N100 IMU from Lunqu were used to collect data. LIO-SAM was used to construct a global 3D point cloud map of the underground parking lot. Figure 2 Segment the ground points, use PCA to estimate the ground plane normal vector, rotate the point cloud map so that the plumb direction of the point cloud map is parallel to the plumb direction of the world coordinate system, and the result is as follows Figure 3-1 , Figure 3-2 As shown in the figure. Relying on the octree data structure, the global point cloud can be quickly denoised. With PCA, the wall column point cloud is roughly segmented; based on the z-axis information of the point cloud, a histogram is constructed to observe the z-value distribution of the point cloud, select a specific threshold, and extract the point cloud slice. Remove the ground points and ceiling points; based on the clustering algorithm, further cluster the wall column point clusters, remove the vehicle point cloud, and project it to the XOY plane. The result is as follows: Figure 4 shown.
[0078] Based on the discreteness of each point cluster, the column point cluster and the wall point cluster are preliminarily distinguished, and the column point cluster is sequentially subjected to secondary clustering segmentation based on DBSCAN and short-line wall point cluster screening based on PCA to obtain the final column point cluster and wall point cluster. The results are shown in the figure. Figure 5 , Figure 6 shown.
[0079] For the column point cluster, RANSAC is used to fit two mutually perpendicular straight line segments, and further construct the four sides of the square and solve its four vertices. For the short-line wall point cluster, the straight line segment is directly fitted by the RANSAC algorithm. For the non-linear wall point cluster, the binary image of the point cloud is first constructed, and then the expansion, corrosion and thinning operations in image processing are performed in sequence to fit the wall line. After fitting, the wall line is smoothed by Gaussian filtering, and the result is as follows: Figure 7 Finally, all the fitted wall and column curves are drawn together, as shown in Figure 8 shown.
[0080] The effectiveness of the proposed method is verified by three sets of experiments. The first set of experiments verifies its accuracy by comparing the results obtained by the experimental method with the actual ground contour. In this experiment, the original CAD file is analyzed and the plane structure information of the corresponding wall columns is extracted as the real ground contour information, such as Fig.10 and Fig.11 The results show that the fitted wall column curve is basically consistent with the real ground contour information. However, Fig. 9 The fitted curve in the red dashed box should not exist in the actual parking lot ground contour. This is because during the data collection process, the LiDAR light scans the wall behind the open door or transparent glass, resulting in the acquisition of wall point cloud information in these areas. This is a factor that is difficult to avoid in the actual collection process.
[0081] The second set of experiments evaluates the performance of this method by comparing the number of columns extracted by this method with the actual number of columns. In the actual parking environment, the orientation of the columns and the columns connected to the wall have little effect on pedestrian or vehicle navigation. Therefore, in this experiment, only the number of columns not connected to the wall is counted, and their types are as follows: Fig.11 The red columns in the middle are shown. The statistical results show that the number of columns extracted by this patented method is 89, with an accuracy rate of 97.75% and a recall rate of 95.60%. The columns that are extracted by mistake are mainly due to the fact that the LiDAR light penetrates the glass during the acquisition process, resulting in the column information in the outer area of the parking lot. The columns that are missed are mainly due to the fact that only one surface of the column is scanned during the acquisition process, making it difficult to distinguish whether it is a column or a short-line wall.
[0082] The third group of experiments measures the wall line extraction accuracy of the patented method by using the area of the individual regions formed after the wall line extraction. CloudCompare software is used to calculate the area of the same region in the constructed three-dimensional point cloud map as the approximate true value. Fig.12 The three closed areas shown in the figure are used to calculate the area of the closed area enclosed by the wall line points extracted by the patented method, and the results are shown in Table 1. The relative errors of the three areas are all within a reasonable range, indicating that the patented method has high accuracy and reliability in wall line extraction. Table 1 Accuracy evaluation of wall line extraction results
Claims
1. A method for extracting ground contours of underground parking lots based on laser radar 3D point cloud, characterized in that: The ground contour extraction method includes a point cloud information preprocessing module, a wall column segmentation and classification module, and a curve fitting module. Compared with the conventional ground contour extraction direction, this method can effectively filter the noise point cloud, and solve the problem of wall column point cloud segmentation in the parking lot environment through point cloud slicing segmentation and adaptive clustering algorithm. By effectively classifying the wall column point clusters and designing the corresponding curve fitting algorithm, this method can achieve accurate ground contour extraction in a parking lot environment with severe occlusion without presetting the number of straight line fittings.
2. The point cloud information preprocessing module according to claim 1, characterized in that: Based on the real-time positioning and mapping technology of LiDAR-IMU fusion, the collected LiDAR keyframe point clouds are aligned and spliced to construct a global 3D point cloud map of the underground parking lot. The ground point cloud in each LiDAR keyframe is then segmented by the ground segmentation algorithm to generate a global ground plane point cloud. The iterative principal component analysis (PCA) method is used to estimate the eigenvector of the ground plane and calculate its angle with the unit z-axis vector. The corresponding rotation matrix is solved to rotate the global 3D point cloud accordingly to achieve the plumb direction correction of the global point cloud map. Finally, the distance-based and PCA-based denoising algorithms are performed in turn to obtain a preliminarily processed 3D wall column point cloud.
3. The wall column segmentation and classification module according to claim 1, characterized in that: Construct a point cloud density histogram in the z-axis direction, observe its distribution, and cut point cloud slice data of a certain thickness to remove ground points and ceiling points. Use the DBSCAN clustering algorithm to perform preliminary clustering and data dimensionality reduction on the wall and column point clouds in the parking lot, project the point clusters onto the XOY plane, and obtain a two-dimensional wall and column point cloud. Based on the discreteness of the point cloud, the preliminary classification of the wall column point clusters is realized. In order to solve the problem that some columns are too close to each other, resulting in incomplete point cluster classification, the DBSCAN clustering algorithm is used to perform secondary clustering segmentation on each initially classified column point cluster. In order to solve the problem that the short straight-line wall point clusters are difficult to distinguish from the column point clusters due to their small discreteness, the PCA algorithm is used to analyze the point distribution of the column point clusters after secondary clustering segmentation, and the linearly distributed point clusters are reclassified as linear wall point clusters.
4. The curve fitting module according to claim 1, characterized in that: This module includes two parts: column curve extraction and wall curve extraction. Column curve extraction traverses each column point cluster, fits two mutually perpendicular straight line segments based on the random sampling consensus (RANSAC) algorithm, and then constructs the four sides of the square and solves its vertices. The column shape and coordinate information are stored in the column information structure and drawn. Wall curve extraction uses the RANSAC algorithm to directly fit each short-line wall point cluster; for non-short-line wall point clusters, a binary point cloud image is generated, and through expansion, corrosion and Zhang-Suen thinning operations, the refined highlight pixel coordinates are finally connected, and the Gaussian filter is applied to smooth the curve. Finally, the smoothed wall line is divided into point clusters and stored in the wall information structure and drawn.
Citation Information
Cited By
Method and device for generating outdoor large-scale scene laser radar point cloud map
CN119810360A
Method and device for generating outdoor large-scale scene lidar point cloud map
CN119810360B
Unmanned bus station accurate parking method and system based on vehicle-road cooperation
CN120340295A
