Research On The Generation Method Of Trees And Vehicles 3D Point Cloud In Autonomous Vehicle Virtual Testing

Autonomous vehicle virtual testing meets the higher testing requirements of autonomous vehicles and has become an important part of the testing of autonomous vehicles.The 3D point cloud generated by Li DAR is an important data source for autonomous vehicle to perceive the surrounding environment.Traffic elements such as pedestrians,vehicles,roads,and trees are an important component of traffic scenes.Therefore,the simulation and generation of the 3D Li DAR point cloud of the above-mentioned traffic elements is one of the important tasks of autonomous vehicle virtual testing.This paper uses two methods of point cloud generation based on object surface models and data-driven to generate point cloud for trees and vehicles respectively,which are two important elements of traffic scenes.The main work of this paper is as follows:Ⅰ.A fast tree point cloud generation method based on the space transformation of the billboard is proposed.According to the texture image of the trees,a two-dimensional plane point cloud of trees is obtained,and the final three-dimensional point cloud of trees is obtained after contour extraction,homogenization,spatial rotation,random offset and scale prior transformation.A method for evaluating the similarity of the three-dimensional point cloud of the spatial histogram is proposed.The three-dimensional point cloud space is projected into several subspaces,and the Bhattacharyya coefficient is used to calculate the similarity of the spatial histogram of each projection subspace.The weighted similarity of projection spatial histogram is used as the evaluation of the point cloud similarity.The experimental results show that the point cloud generated by the method in this paper can express the geometric shape of trees,and the average similarity with the point cloud generated by the geometric model is more than 90%,while the point cloud generation time is only 1% of the geometric model method.And the similarity evaluation method of point cloud in this paper can effectively evaluate the similarity between point clouds.Ⅱ.A two-stage dense point cloud generation framework based on a single view is proposed.This framework consists of the two-branch decoder sparse point cloud generation network and the multi-layer perceptron dense point cloud generation network.An autoencoder structure is applied in the sparse point cloud generation network.The encoder extracts the multichannel features of the single-view vehicle image through the stack of multi-layer convolutions,and the decoder maps the features into a three-dimensional point cloud.The method of Min-ofN loss is used to reconstruct the diversity of point cloud,and finally the most reasonable point cloud is generated.In the dense point cloud generation network,this paper adopts the idea of Point Net++ hierarchical structure.The local and global feature of the point cloud are extracted through shared multi-layer perceptron,which meets the requirements of point cloud feature input order invariance and spatial transformation invariance,and the unity of local and global.Combining the characteristics of EM distance and Chamfer distance,and a two-stage training method is adopted to generate dense point cloud in different densities.The experimental results show that the two-stage dense point cloud generation framework can generate high quality dense point cloud of vehicles.EM distance and Chamfer distance are 38% and 37% lower than single-stage dense point cloud generation networks respectively,and the average spatial histogram similarity is 2.75% higher than single-stage dense point cloud generation networks.