Research And Application Of Cloud Segmentation Method For Large Scenic Spots Based On Graph Convolution

In recent years,with the development of space perception and large-scale processing technology of point cloud,laser-based 3D scanning technology has been widely used in real-time navigation,virtual reality,building information model and other fields.For these applications,point cloud segmentation technology is one of the necessary technologies.Point cloud segmentation is the key technology to effectively utilize the point cloud model and it achieves the separate treatment of the point cloud model.Point cloud segmentation technology is necessary to segment and extract models whether it is utilized to recognition or classification.It is main research area to improve accuracy and efficiency for point cloud segmentation.This thesis proposes a graph convolution neural network,and improves the method of point cloud segmentation by using the optimization method of label data and semantic instance segmentation.The method is improved to solve the problems of large scale and uneven distribution of point cloud in the cloud data of large scenic spots.The accuracy of point cloud segmentation is improved,and the memory consumption and running time in training are reduced.The main research work includes:(1)A method of point cloud feature extraction based on graph convolution neural network is proposed.Aiming at the characteristics of the point cloud in the large scenic spots without specific sequence and uneven distribution,the distance information of the sampling points is sorted by using the neighborhood sampling method based on k-adjacent sampling method,and the Laplace matrix of the input point cloud is calculated.The sampling method is improved by referring to the hollow convolution technology to expand the network sensing area.In view of the problem that the graph convolution is easy to produce gradient disappearance,the residual cloud is added in this thesis to improve the ability of network resisting over smooth.(2)A point cloud label optimization method based on label propagation is proposed.In order to solve the problem of deep learning demand for large data,this thesis optimize instance label and semantic label.The optimized label matrix and the tag label matrix of the original data set are used to influence the result of point cloud segmentation.(3)An improved method of point cloud segmentation is proposed.Aiming at the similarity between semantic label and instance label of point cloud data,the similarity parameters are combined into the loss function of semantic and instance segmentation.At the same time,the segmentation results are affected by adjusting the proportion of similarity parameters to process semantic segmentation and instance segmentation of point cloud.Through the experiment of point cloud segmentation in open data set S3 DIS,compared with the segmentation results of point net and point net++,it is proved that the combination of semantic segmentation and instance segmentation can improve the efficiency and accuracy of point cloud segmentation.