Tunnel/roadway Point Cloud Registration And Segmentation Technology Based On Deep Learning And Its Application

3D laser scanning has the characteristics of non-contact,fast,high-precision,and large-scale,and has become an important tool for measurement and analysis in underground space.Point cloud registration is an important step in obtaining complete area point clouds in complex underground spaces.Studying point cloud segmentation algorithms for tunnels/tunnels in underground space scenes is of great significance for information extraction in this scene.This article is based on deep learning technology and focuses on point cloud registration and point cloud segmentation algorithms in underground tunnel/roadway scenes.The main research content is as follows:(1)A new point cloud registration network model,DGRNet,was constructed based on PCRNet and combined with the advantages of edge convolutional networks in local features.In the feature extraction module of the network,edge convolution is used to check the input point cloud for feature learning,which can better learn the complex feature changes and geometric structure of the 3D point cloud and improve the understanding ability of local features in the scene.The experimental results of object models and underground tunnel scenes show that the DGRNet network has better registration accuracy,and its registration accuracy is stable under the influence of point cloud noise,with good robustness.(2)A point cloud segmentation network model KNPNet based on randomly expanding K-neighborhoods was constructed to address the issue of lack of local information in Point Net.The network extracts local features and point neighborhood relationships by randomly expanding the K-neighborhood,and uses the point pyramid pooling layer to induce local features.It is combined with the maximum pooling layer to enhance the understanding ability of local features in the scene.The experimental results show that the KNPNet network has better segmentation performance than existing mainstream methods in component segmentation and underground tunnel point cloud scenes.(3)A set of point cloud datasets based on measured tunnels was created.Rotate and correct the point cloud data of the actual tunnel,parallel to the X-axis,and manually remove noise points after comparing with the actual situation.Create label fields to assign different values to different categories in the point cloud,providing label basis for semantic segmentation of the point cloud.The annotated point cloud data is divided along the X-axis direction with the same length,achieving the production of tunnel datasets.(4)Developed a point cloud visualization software interface.Based on the Visual Studio cross platform compilation environment and Qt Creator interface,functions such as point cloud reading,point cloud display,point cloud filtering and noise reduction,point cloud downsampling,and point cloud 3D reconstruction have been implemented.There are 44 pictures,7 tables and 94 references in this thesis.