关键词: 自动驾驶汽车虚拟测试, 三维点云, 树木, 布告牌, 空间直方图

Abstract:

Autonomous vehicle virtual testing has become an important method for autonomous driving or vehicle-road collaborative testing and evaluation. Virtual generation of LiDAR data is one of the most important tasks in autonomous vehicle virtual testing. At present, the geometric model method based on the time-of-flight principle is a common method to generate LiDAR 3D point cloud data, which may lead to poor performance for real-time generating point cloud. The billboard technology is often used to model trees in virtual scenes. Since the billboard is only composed of two rectangular meshes with eight triangles, the 3D point cloud data generated by the billboard directly can hardly reflect the real spatial information of trees. To solve the above problems, this paper proposes a fast 3D tree point cloud generation method based on the spatial transformation of the billboard. Based on the texture image of the billboard, the 2D point cloud of trees is obtained according to the texture transparency, and the boundary of the 2D tree point cloud is estimated. Combined with the prior knowledge of the tree structure, the rotation, random offset and scale transformation are carried out in turn. After the above operation, the 3D point cloud data of trees can be obtained with less triangle meshes and less computational cost. This paper also proposes a similarity evaluation method of the 3D point cloud based on spatial histogram. The 3D point cloud space is quantified into several subspaces, and the projection spatial histograms of the 3D point cloud are obtained, and the Bhattacharyya coefficient is adopted to calculate the similarity of two spatial histograms. The weighted similarity of projection spatial histogram is used as the evaluation of the point cloud similarity. The experimental results show that the average similarity between the 3D point cloud data of spruce and other three trees generated by the proposed method and by the geometric modeling methods is more than 90%. Compared with the geometric modeling methods, the method only takes about one percent of the time to generate 3D tree point cloud. The 3D point cloud of trees generation method based on the spatial transformation of the billboard is fast and accurate and can meet the performance requirements of autonomous vehicle virtual testing.

Key words: autonomous vehicle virtual testing, 3D point cloud, tree, billboard, spatial histogram