A Framework For 3D Point Clouds Quality Evaluation And Completion Of Urban Vehicles

In recent years,autonomous driving technology has gradually become the focus of research in academia and industry,and the 3D point cloud data obtained by vehicle-borne LiDAR is the key information for intelligent vehicles to perceive environment.Due to the relative position of the LiDAR sensor and the object,the self-occlusion of object and the mutual occlusion between objects,the system cannot obtain a complete point cloud of the vehicle during data collection.To perceive the scene accurately,it is necessary to get complete and high-quality point cloud data through 3D completion method.With the development of related research,many deep learning-based 3D completion models have emerged one after another,but most of these studies aim at synthetic data,and the accuracy will decrease when they are transferred to real data.Further improving the accuracy and credibility of the model is of great significance for realizing automatic driving in complex road scenes.In response to this problem,this dissertation proposes an end-to-end 3D point cloud quality evaluation and completion model(Point Voxel Completion Network,PVCNet)for vehicle,the main contents are as follows:1.We build a simulated radar scan data set of vehicles.By analyzing the characteristics of the point cloud data of vehicles in the real road scene,using the method of simulated radar scanning,the simulation data is automatically generated,and sufficient training data is provided for the proposed deep learning framework PVCNet.2.We design a point cloud data quality evaluation system.Focusing on the inconsistency of point cloud density and large differences in the degree of occlusion,an automatic point cloud quality evaluation system is designed,combined with the shape and density of the point cloud,to evaluate the quality of the point cloud,and provide a confidence reference for the point cloud completion result.At the same time,it preliminarily predicts the location information of the missing point cloud and transfers it to the final completion module in the form of feature vectors.3.We propose a point cloud completion model for road vehicles.According to the characteristics of the point cloud data of vehicles in the road scene,the normal vector loss and the main direction loss are designed based on the 3D reconstruction loss.Moreover,the advantages of point cloud data and voxel data are combined to improve the points for vehicles in the road scene.Through a number of experiments in different scenarios and different tasks,it is proved that the PVCNet framework proposed in this paper has good effects on the task of 3D point cloud completion for road vehicles,and it also demonstrates a competitive performance on the task of completing multi-category point cloud objects.