Research On 3D Object Detection Algorithm Based On Point Clou

Autonomous driving technology is an important component of future intelligent transportation systems,and three-dimensional object detection is one of the core technologies in autonomous driving.3D target detection aims to extract the position,size,shape and other information of objects from 3D point cloud data of laser radar,so as to help the auto drive system recognize and predict objects in the surrounding environment,which is of great significance to improve the safety and reliability of autopilot.This thesis proposes two 3D object detection algorithms based on the sparsity of Li DAR point clouds using deep learning methods,and conducts theoretical analysis,method validation,and result analysis.The main research content is as follows:(1)In response to the differences in target feature distribution caused by different sparsity of 3D point cloud targets in existing Li DARs,as well as the insufficient flexibility of existing algorithm feature sampling strategies,this thesis proposes a 3D object detection algorithm based on deformable voxel feature pyramid.Firstly,an adaptive voxel feature pyramid module was proposed,which selects the corresponding feature layer for feature aggregation based on the sparsity of non empty voxels of the region of interest(Ro I).Secondly,a deformable voxel Ro I pooling operation is proposed,which extracts non empty voxel dense regions through deformable grid points to enrich semantic features and gather relevant background information outside of Ro I to improve positioning accuracy.The experiment shows that in the widely used KITTI dataset vehicle classification,our method outperforms the 3D object detection method PV-RCNN in terms of average precision(AP),which increases by 0.47%,1.63%,and 1.34% in three levels: simple,medium,and difficult,respectively.(2)To address the issue of insufficient single stream feature information in single point cloud or single voxel 3D object detection algorithms,this thesis proposes a 3D object detection algorithm based on multi stream feature aggregation,which utilizes point cloud representations from different perspectives for feature complementarity to achieve accurate 3D object detection.Firstly,multi-stream features are extracted from the point cloud through three representations: original points,voxels,and perspective images.Then,a multi stream feature fusion module based on channel attention was proposed,which effectively utilizes the correlation between different features and eliminates redundant information.The experiment shows that in the widely used KITTI dataset,the proposed method outperforms its baseline method H23D-RCNN in terms of rider classification,with an AP improvement of 5.56%,4.73%,and 5.16% in simple,medium,and difficult levels,respectively.