Research On 3D Point Cloud Object Detection Algorithm Based On Voxel

Due to the widespread use in areas such as autonomous driving and robotics,the research on 3D point cloud object detection has important research value and application prospects.Considering the complex characteristics of 3D point cloud data,this thesis first divides the point cloud into regular voxels,and improves the existing object detection algorithm from different directions based on this.We propose a three-dimensional object detection algorithm framework based on path augmentation——3D BFP-RPN Net,3D GIoU Net algorithm that focuses on improving the detection boxes positioning,and design a general voxel feature extraction module VFE++.The key contributions of this thesis are as follows:(1)Based on path augmentation and feature fusion weight self-learning,a voxelbased 3D object detection network framework——3D BFP-RPN Net is designed.The architecture fully learns point cloud features through two-way information channels,and the organic fusion of features at various levels enhances the network’s ability to recognize objects of multiple sizes.Through the comparison of experimental results on the public dataset,it is proved that this detection algorithm has a stronger detection ability than the algorithm before the improvement,especially for the detection of small samples and difficult samples.(2)In order to improve the positioning quality of the detection boxes and make up for the loss of the detection boxes positioning score in the detection process,a network framework including a positioning score prediction module——3D GIoU Net is designed.The network corrects the position of the detection box by predicting the 3D Generalized Intersection over Union(3D GIoU)between the predicted boxes and the true boxes.In addition,in non-maximum suppression,the 3D GIoU value of the detection boxes are regarded as the positioning score to assist in screening and removing duplicate detection boxes.On the public dataset,the detection accuracy of the 3D GIoU Net is significantly higher than that of the benchmark algorithm.(3)The existing general voxel feature extraction module ignores the internal features of voxels and cannot capture fine details.For this reason,an improved voxel feature extraction module——VFE++ module has been designed.The VFE++ module can capture different levels of local spatial features within a voxel,thereby helping the network better learn point clouds and detect objects.The module is not limited to a special network structure,and can be extended to any voxel-based 3D object detection algorithm.By porting this module to the previously proposed 3D BFP-RPN Net and 3D GIoU Net the accuracy of the corresponding algorithm on the public dataset is further improved.