Research On Point Cloud Scene Semantic Segmentation Based On Deep Learning

At present,more and more intelligent application scenarios are inseparable from the semantic segmentation technology of three-dimensional scenes.For example,in an unmanned environment,the ability to perceive the surrounding environment during the travel of a car is conducive to improving its positioning or obstacle avoidance.In order to realize the method of point cloud scene semantic segmentation based on deep learning,this paper has learned the relevant theories and knowledge of laser point cloud scene data,deeply studied the theory of deep neural network and point cloud-based scene segmentation network Point Net,and concluded that the network is defective in local feature extraction and propose corresponding improvement schemes.Experiments show that the improved method proposed in this paper improves the accuracy of the original network in semantic segmentation tasks.The specific work done in this article is as follows:1.Aiming at the problem of insufficient local point cloud feature extraction capability for Point Net networks.Inspired by SIFT feature descriptors in 2D images,this paper designs a multi-scale neighborhood awareness module built by the 3D point cloud feature description network Point SIFT.This module eliminates the threshold debugging process in traditional feature operators,Selflearning to local characteristics of point cloud scenes with direction perception and scale adaptation.This article uses this as a basis to improve the Point Net network,that is,the local features extracted by Point SIFT and the output features of the Point Net network are fully connected multiple times,and finally obtain the classification score of each point to achieve semantic segmentation.2.Aiming at the problem of semantic segmentation of outdoor point cloud large scenes.This paper designs a set of experimental schemes for data collection,processing,scene production,and semantic segmentation of point cloud scenes.First,a ground laser scanner is used to multi-sitely collect the area near the school library,and the original point cloud data preprocessing such as streamlining and denoising,from which buildings with more overlapping areas in each two-view scene are selected,and an automatic registration method based on Sample Consensus Initial Alignment and Iterative Closest Point is used to obtain point cloud data for large outdoor scenes.Then use the Semantic3 D data set as a sample to train an outdoor point cloud scene segmentation model based on Point Net ++.Finally,the model is called to perform semantic segmentation on the reconstructed scene,and the results are visually analyzed.