Research On Key Technologies Of 3D Modeling And Semantic Segmentation For Indoor And Outdoor Scenes

3D scene models are the data base of 3D geographic information service(GIS),building information model(BIM)and virtual reality(VR).It's a universal demand of urban management,planning and intelligent transportation system to create and update accurate urban spatial model.In the past decades,with the comprehensive popularization of digitalization and the rapid upgrading of communication bandwidth,there are more and more demands on 3D information services.Therefore,it is of great significance to study the automatic 3D scene semantic modeling technology.This dissertation focuses on the automatic semantic modeling of indoor and outdoor3 D scenes.Fast acquisition and semantic segmentation of 3D point cloud are the basis of3 D semantic modeling of indoor and outdoor scenes.However,according to the current research status,the mobile laser point cloud acquisition technology in the environment without GNSS is not enough,and the point cloud scene understanding technology is still in the initial research stage.Therefore,we make a detailed study on point cloud data acquisition and semantic segmentation of 3D scene,and lays the foundation for automated semantic 3D scene modeling.The main works of this dissertation are as follows:1.3D point cloud data acquisition of indoor and outdoor scene based on simultaneous localization and mapping(SLAM).The technology of mobile laser point cloud acquisition is not mature in the absence of GNSS.Firstly,three 2D laser scanners are used as the core machines to design the hardware platform.Then the time synchronization and position and attitude calibration for sensors are performed.Finally,the laser odometer based on particle filter and the loop detection based on multi-resolution submaps are used to realize the SLAM,and the 3D point cloud data of the scene is obtained.The experimental results show that the accuracy of the 3D point cloud data acquisition system proposed can reach 6cm,which can provide reliable experimental data for point cloud scene segmentation and understanding.2.The density effect of point cloud in indoor and outdoor point cloud scene segmentation.Based on the fact that the density of point cloud is limited by the ability of scanning equipment,different applications have different requirements for density,and point cloud sampling is a conventional preprocessing.First,we analyze the density effect in the conventional point cloud data processing.Then,the analysis and research method of point cloud density effect is designed for deep learning.Finally,the density effect in point cloud scene segmentation using deep learning is summarized from the influence on target extraction and the generalization ability of neural networks.This work provides a reference of point cloud density for indoor and outdoor data collection and neural network design and training.3.The scene segmentation of dense indoor point cloud based on irregular sampling coding and voxel 3D convolution.For the point cloud data does not have regular adjacency,there are large scale differences among the point cloud targets,and the distribution of indoor point cloud scene targets is complex and other challenges.Firstly,the idea of indoor dense point cloud segmentation based on spatial sampling and local feature encoding is proposed.Then,the spatial sampling method and local feature coding method of point cloud data are studied.Finally,a full convolution point network with scattered encoders in high-resolution space and voxel convolution in low-resolution space(HSLV-FCPN)is designed to segment dense indoor point cloud scenes with rich detail and fast association feature extraction.Experiments on Scan Net data set show that it has a 5% improvement over FCPN in some categories' Io U.It means that the work can realize the segmentation and extraction of indoor objects,and promote the automatic semantic modeling of indoor3 D scene.4.A two-level segmentation structure for street level point cloud scene based on energy local clustering and gated graph neural network global classification.Due to the problems of large spatial scale and large amount of data in the street level point cloud scene,there are few researches on the street level mobile laser point cloud scene segmentation network.Firstly,the two-level segmentation method is proposed.Then,the graph cutting method based on energy function is used to cluster the point cloud into point clusters of different sizes unsupervised.Finally,the point clusters are organized into graph structure by adjacency relationship,the initialization module of graph state is completed by point cloud feature extraction in the cluster,the update module of graph state is completed by convolution operation of adjacency information between clusters.The gated graph neural network for point clusters classfiction(GGNN4PC)is constructed to segment the street level point cloud scene fastly.The experiment on the vehicle mobile measurement data sets show that the m Io U is 5% higher than SPG.It can achieve the target segmentation and extraction of the street scene with relatively clear structure,but it still has some difficulties in the scene surrounded by vegetation.This is of great significance for the rapid and automatic creation and updating of accurate urban street models.