CN-121982131-A - Point cloud-based land line generation method and device and intelligent driving equipment

CN121982131ACN 121982131 ACN121982131 ACN 121982131ACN-121982131-A

Abstract

The invention discloses a point cloud-based land line generation method, a point cloud-based land line generation device and intelligent driving equipment, and relates to the technical field of point cloud processing. And splicing the plurality of original point clouds under a vehicle coordinate system through external parameter transformation to obtain fusion point clouds. Inputting the fusion point cloud into a pre-trained point cloud segmentation neural network, performing semantic segmentation processing on the fusion point cloud by the point cloud segmentation neural network, and outputting a two-dimensional grid graph. And carrying out grid index mapping on each three-dimensional point based on the two-dimensional grid graph, and obtaining a non-ground point set. And generating a grid index sequence by taking the projection of the origin of the coordinate system of the corresponding laser radar on the two-dimensional grid graph as a starting point and taking the projection of each non-ground point on the two-dimensional grid graph as an ending point. And selecting land line points from non-land points of all grids of the grid index sequence, and generating a land line according to a plurality of the land line points.

Inventors

WU JIN
HU DONGHONG
WANG JIALIN
SHU LEIZHI
GUO XINYANG
WU ENGUANG

Assignees

城市之光(深圳)无人驾驶有限公司

Dates

Publication Date: 20260505
Application Date: 20260408

Claims (10)

1. The ground line generation method based on the point cloud is characterized by comprising the following steps of: Acquiring original point clouds acquired by a plurality of laser radars of a vehicle; splicing a plurality of original point clouds under a vehicle coordinate system through external parameter transformation to obtain fusion point clouds; inputting the fusion point cloud into a pre-trained point cloud segmentation neural network, performing semantic segmentation processing on the fusion point cloud by the point cloud segmentation neural network, dividing a non-ground area from a three-dimensional space, and outputting a two-dimensional grid map, wherein each grid of the two-dimensional grid map is stored with a classification label which is used for identifying whether a space area corresponding to the grid is the non-ground area or not; Grid index mapping is carried out on each three-dimensional point in the fusion point cloud based on the two-dimensional grid graph, and non-ground points positioned in a non-ground area are screened out from the three-dimensional points according to the classification labels, so that a non-ground point set is obtained; Taking the projection of the origin of a coordinate system of a corresponding laser radar on a two-dimensional raster image as a starting point and taking the projection of each non-ground point in the non-ground point set on the two-dimensional raster image as an end point to generate a raster index sequence from the starting point to the end point; and selecting land line points from non-land points of all grids of the grid index sequence, and generating a land line according to a plurality of the land line points.
2. The land line generating method based on the point cloud as claimed in claim 1, wherein grid index mapping is performed on each three-dimensional point in the fused point cloud, specifically: For three-dimensional points Calculating the corresponding grid index according to the following formula ; ; In the formula, Is the coordinates of the three-dimensional point in the vehicle coordinate system, Is the minimum coordinate value of the rectangular area of the two-dimensional grid graph in the Y-axis direction and the X-axis direction, Spatial resolution of the two-dimensional raster pattern; Representing a rounding down.
3. The method for generating a point cloud based ground line according to claim 1, wherein generating the grid index sequence from the start point to the end point comprises the steps of: converting the projection coordinates of the starting point and the end point in the two-dimensional grid graph into grid indexes respectively Grid index ; Generation of slave grid indices using two-dimensional Bresenham algorithm To grid index All grid indexes passed through; Generating grid index sequence by all grid indexes in sequence Wherein n is a positive integer, 。
4. The method for generating a ground line based on a point cloud as recited in claim 1, wherein the ground line points are selected from non-ground points of all grids of the grid index sequence, comprising the steps of: Sequentially inquiring classification labels of all grids along the grid index sequence; if the classification label of the grid is marked as a non-ground area, outputting a non-ground point of the grid as a candidate point; and calculating Euclidean distances between all candidate points of the grid index sequence and the origin of a coordinate system of the laser radar, and outputting the candidate point with the minimum Euclidean distance as the ground line point.
5. The point cloud based ground line generation method of claim 4, wherein: And outputting the non-ground point corresponding to the end point as a ground line point if the classification labels of all grids of the grid index sequence identify that the non-ground area does not exist.
6. The land line generating method based on the point cloud as claimed in claim 4, wherein a non-ground point of the grid is outputted as a candidate point, concretely: When the grid comprises only one non-ground point, outputting the non-ground point as a candidate point; When a plurality of non-ground points exist in the same grid, the Euclidean distance between each non-ground point and the origin of the coordinate system of the laser radar is calculated, and the non-ground point with the minimum Euclidean distance is used as a candidate point to be output.
7. The ground line generation method based on point cloud according to any one of claims 1 to 6, wherein inputting the fused point cloud into a pre-trained point cloud segmentation neural network further comprises preprocessing the fused point cloud to obtain a network input feature tensor, and specifically comprises the following steps: performing height clipping on the fusion point cloud, and reserving three-dimensional points in a preset height range; dividing the reserved three-dimensional points according to the preset voxel resolution, and counting the points in each voxel; cutting off the number of three-dimensional points in the voxels, and calculating a logarithmic value to obtain the feature of the number of the voxels; extracting the voxel statistical characteristics of each voxel, and splicing to form a 30-dimensional voxel characteristic vector; Dividing the reserved three-dimensional points in an XY plane according to Pillar resolution, and extracting the maximum height and the minimum height based on the three-dimensional points in each Pillar to form 2-dimensional geometric features; carrying out element-by-element maximum pooling on 30-dimensional voxel feature vectors of a plurality of voxels covered by each Pillar to obtain aggregated voxel features, and splicing the aggregated voxel features with corresponding 2-dimensional geometric features to form 32-dimensional feature vectors; arranging the 32-dimensional feature vectors of Pillar according to the corresponding grid coordinates to form a three-dimensional feature tensor; generating a binary Confidence Mask with the same size as the three-dimensional feature tensor, setting a Pillar position containing three-dimensional points to be 1, setting a Pillar position not containing three-dimensional points to be 0, and splicing the binary Confidence Mask serving as an additional channel with the 32-dimensional feature vector to form the network input feature tensor.
8. A point cloud-based ground line generation device, comprising: The acquisition module is used for acquiring original point clouds acquired by a plurality of laser radars of the vehicle; The splicing module is used for splicing the plurality of original point clouds under a vehicle coordinate system through external parameter transformation to obtain fusion point clouds; The segmentation module is used for inputting the fusion point cloud into a pre-trained point cloud segmentation neural network, performing semantic segmentation processing on the fusion point cloud by the point cloud segmentation neural network, dividing a non-ground area from a three-dimensional space and outputting a two-dimensional grid graph, wherein each grid of the two-dimensional grid graph is stored with a classification label, and the classification label is used for identifying whether a space area corresponding to the grid is the non-ground area or not; The screening module is used for carrying out grid index mapping on each three-dimensional point in the fusion point cloud based on the two-dimensional grid graph, and screening non-ground points positioned in a non-ground area from the three-dimensional points according to the classification labels to obtain a non-ground point set; The index module is used for generating a grid index sequence from the starting point to the end point by taking the projection of the origin of the coordinate system of the corresponding laser radar on the two-dimensional grid graph as the starting point and taking the projection of each non-ground point in the non-ground point set on the two-dimensional grid graph as the end point; and the generating module is used for selecting ground line points from non-ground points of all grids of the grid index sequence and generating ground lines according to a plurality of ground line points.
9. An electronic device comprising a processor coupled to a memory, the memory storing program instructions that when executed by the processor implement a point cloud based ground line generation method of any one of claims 1 to 7.
10. An intelligent driving apparatus comprising the point cloud-based ground line generation device according to claim 8.

Description

Point cloud-based land line generation method and device and intelligent driving equipment Technical Field The invention relates to the technical field of point cloud processing, in particular to a point cloud-based land line generation method and device and intelligent driving equipment. Background In a low-speed scenario of an autopilot system (such as a parking lot, a park road or a narrow channel, etc.), a ground line is generally used to indicate a passable area of a vehicle or road boundary information, and thus, accurately detecting the ground line and acquiring a spatial position thereof is of great significance in providing travel constraint data for a vehicle planner. Currently, the existing land line detection methods in low-speed scenes mainly comprise two types. One type is a method based on visual segmentation, which comprises the steps of acquiring an environment image through a vehicle-mounted camera, identifying a ground line area by using a deep learning network, and recovering the spatial position of the ground line area through a two-dimensional to three-dimensional projection method. The other type is a method for constructing a grid map based on a laser radar point cloud, wherein the laser radar source point cloud is constructed as a grid map, and the ground line characteristics are extracted from a grid space. However, both of the above methods still have certain limitations. The vision-based method relies on conversion from a two-dimensional image to a three-dimensional space, so that the space positioning accuracy is easy to be obviously reduced when a scene with gradient change or a ground line is far away from the optical center of the camera. The method for constructing the grid map based on the laser radar is limited by the design of the grid resolution, and generally only the detection precision of the grid scale can be achieved, so that the ground line detection precision is limited. Therefore, how to improve the detection accuracy of the ground line in the low-speed scene is still a technical problem to be solved in the art. Disclosure of Invention In order to solve the technical problems that in the prior art, the positioning accuracy of a space based on a two-dimensional image is low and the land line detection accuracy is insufficient due to the fact that a grid resolution based grid mapping method is limited, the invention aims to provide a land line generation method and a land line generation device based on point clouds. The aim of the invention is achieved by the following technical scheme: In a first aspect, the present invention provides a point cloud-based land line generating method, including the steps of: Acquiring original point clouds acquired by a plurality of laser radars of a vehicle; splicing a plurality of original point clouds under a vehicle coordinate system through external parameter transformation to obtain fusion point clouds; inputting the fusion point cloud into a pre-trained point cloud segmentation neural network, performing semantic segmentation processing on the fusion point cloud by the point cloud segmentation neural network, dividing a non-ground area from a three-dimensional space, and outputting a two-dimensional grid map, wherein each grid of the two-dimensional grid map is stored with a classification label which is used for identifying whether a space area corresponding to the grid is the non-ground area or not; Grid index mapping is carried out on each three-dimensional point in the fusion point cloud based on the two-dimensional grid graph, and non-ground points positioned in a non-ground area are screened out from the three-dimensional points according to the classification labels, so that a non-ground point set is obtained; Taking the projection of the origin of a coordinate system of a corresponding laser radar on a two-dimensional raster image as a starting point and taking the projection of each non-ground point in the non-ground point set on the two-dimensional raster image as an end point to generate a raster index sequence from the starting point to the end point; and selecting land line points from non-land points of all grids of the grid index sequence, and generating a land line according to a plurality of the land line points. In one possible implementation manner, grid index mapping is performed on each three-dimensional point in the fused point cloud, specifically: For three-dimensional points Calculating the corresponding grid index according to the following formula; ; In the formula,Is the coordinates of the three-dimensional point in the vehicle coordinate system,Is the minimum coordinate value of the rectangular area of the two-dimensional grid graph in the Y-axis direction and the X-axis direction,Spatial resolution of the two-dimensional raster pattern; Representing a rounding down. In one possible implementation manner, the grid index sequence between the starting point and the end point is generated, which spec