Research On PointUnet Network For Semantic Segmentation Of Scattered Point Clouds In Indoor 3D Scenes

The semantic segmentation of scattered point cloud in 3D scenes is the premise and foundation for the realization of indoor 3D scene understanding and intelligent environment perception.In recent years,domestic and foreign scholars have continuously tried to use deep learning for segmentation and recognition of scattered point cloud in indoor 3D scenes.However,the inherent disorder of 3D scattered point cloud data makes the existing scattered point cloud deep learning models still unable to avoid the problem of insufficient feature information extraction ability and poor generalization ability of the network framework.This topic aims to improve the point cloud deep learning network feature extraction ability and network frame generalization ability.Based on the local ordering idea and singular value decomposition theory,we conduct research on the semantic segmentation of scattered point clouds in indoor 3D scenes.The research work is as follows:Firstly,aiming at of the problem that the semantic segmentation model of the scattered point cloud in the indoor 3D scene can not avoid the neighborhood aggregation when processing the point cloud with uneven distribution of sampling points,which easily leads to large error in feature information extraction,we proposed a directed neighborhood search strategy.In this method,the nearest neighbor is combined with the quadratic partition of local space,and the "farthest-nearest" point sampling is carried out in each hexagrams to search for the adjacent points,so as to avoid the problem of insufficient neighborhood clustering and feature characterization in the case of uneven distribution of sampling points;refer to Unet network framework,combine directed neighborhood search with PointNet network,and design a new PointUnet network framework for segmentation of scattered 3D point clouds.Secondly,in order to solve the problem of poor adaptability of the network frame of the semantic segmentation model of scattered 3D indoor point cloud,we proposed a singular value decomposition directed convolution calculation model.The model determines the search range of neighboring points and implements local coordinate conversion by performing singular value decomposition on the local covariance matrix;then,by performing directed neighbor search on the local neighbor points after coordinate transformation to impove the model characterization ability;the rotation invariance of the directed neighborhood based on singular value decomposition is proved;through the ordered expression of the singular value decomposition directed neighborhood,a singular value decomposition directed convolution calculation model is constructed.Apply this convolution model to improve the PointUnet network,and introduce the residual module to establish ResPointUnet scattered point cloud semantic segmentation network.Finally,with ScanNet dataset and S3 DIS dataset as the research subject,carry out the experimental research on the semantic segmentation of scattered point cloud data in indoor 3D scenes.In order to verify the performance of the proposed method in semantic segmentation of scattered point clouds in indoor 3D scenes,we test and analyze the overall accuracy rate(oAc),the average accuracy rate(acA)and the mean Intersection Over Union(mIoU)of PointUnet network and ResPointUet proposed in this paper for point cloud scene segmentation;For prupose of verifying the adaptability of the singular value decomposition directed convolution model proposed in this paper,we carry out convolution model transplantation experiments.