Research On Completion And Segmentation For Point Clouds

With more and more convenient data collection,3D data has become a basic data form in computer vision,automatic driving,robot navigation,earth observation and other fields,and the processing of 3D data has also been widely concerned.In point cloud data processing,completion and segmentation are the basic tasks of point cloud repair and understanding,and there are some unsolved problems.First of all,due to the disorder of point cloud structure,point cloud feature extraction generally uses near point search to realize downsampling,and the sampling speed is too slow.Secondly,due to the sparse and uneven structure of point cloud,the precision and density of key points in point cloud will be affected,reducing the quality of completion and segmentation.In order to improve the quality and efficiency of point cloud feature extraction and overcome the problem of uneven point cloud density,this paper takes completion and segmentation as an example to carry out related research,and carries out visualization tool design.(1)To solve the problem of slow point cloud feature extraction,a new encoder structure is proposed in this paper.By introducing correlation matrix and point cloud-voxel interrevolution,efficient point cloud feature extraction is realized in the form of three-dimensional convolution.(2)In the point cloud completion problem,the idea of classification before diffusion regression is proposed to address the inaccuracy of anchor points,so that anchor points undergo classification and regression,so as to improve the accuracy of diffusion datum;(3)Using the newly designed encoder and decoder,this paper designs a new point cloud completion network and segmentation network,and conducts comparison experiments to prove the effectiveness;(4)From the perspective of visual and design convenience,this paper discusses the quality improvement methods of point cloud rendering and browsing,and proposes the realization methods of point cloud visual rendering and three-dimensional browsing based on Point Flow and Unity3 D respectively.