Research On Semantic Segmentation Of 3D Point Clouds In Large-scale Field Scenes

As a common three-dimensional(3D)data,point cloud has been widely used in photogrammetry,remote sensing,indoor navigation,mobile robot,unmanned driving and other related fields in recent years.With the continuous development of sensor technology,the acquisition methods of 3D point cloud data have become more and more diversified.In the past ten years,a series of public datasets for semantic segmentation of large-scale 3D point clouds have been successively released,and datadriven deep learning networks have also flourished.Although 3D point cloud data acquisition and processing technologies are becoming more and more mature,so far,there are few public datasets and effective deep learning networks dedicated to largescale field scenes.Aiming at the above problems,this paper firstly analyzes the research background and existing problems of 3D point clouds semantic segmentation in large scenes.Taking the self-made field scene dataset as the main research object,based on the point-based deep learning network,the Point Net network and the superpoint graph(SPG)network are improved successively,and the performance of the improved network is analyzed.The specific research contents are as follows:(1)Production of a large-scale real field scene dataset.Aiming at the problem of insufficient 3D point cloud data sets currently used for large-scale real scenes in the field,by using the laboratory DJI Phantom 4 RTK industry-level UAV based on the principle of photogrammetry,the kiwifruit industry base around Yangling,on-campus demonstration site and on-campus accommodation area were photographed,and the point cloud data was reconstructed and manually marked in the later stage.A set of datasets dedicated to the task of 3D point cloud semantic segmentation of large-scale real scenes in the field was produced.(2)Design of lightweight point cloud semantic segmentation network based on improved Point Net.In view of the shortcomings of the basic network of Point Net in local feature extraction,this study proposes a more lightweight LFSeg Net network by analyzing and improving its various modules,which solves the important problem of feature information loss well.And compared the performance with Point Net and other networks on the self-made real fruit scene dataset,the results show that our LFSeg Net network has better performance.(3)Construction of large scene point cloud semantic segmentation network based on improved SPG.Aiming at the problem that the superpoint graph(SPG)network has poor performance in the semantic segmentation task of large scene point cloud data,this paper analyzes and improves its feature extraction,superpoint construction,network training and other parts accordingly.And we research how to build a more efficient network model to improve its semantic segmentation performance.A selfmade field real scene dataset is used for experiments.The experimental results show that our improved network has better performance(+15.76% m Io U)than the original SPG network,and verified on the S3 DIS dataset,compared with the original method,the m Io U improved by 1.1 percentage points.The effectiveness of the improved method in this paper is further verified.