Research On Point Cloud Semantic Segmentation And 3D Reconstruction Method Of Urban Underground Pipe Gallery

The urban underground pipe gallery is an important part of the urban underground space,and its realization of three-dimensional model reconstruction is the current difficulty.Traditional modeling based on two-dimensional plane data has been difficult to intuitively and accurately represent the underground space scene.With the development of three-dimensional laser scanning technology,the three-dimensional model based on laser point cloud data is more beautiful and accurate.Rapid model reconstruction and semantic segmentation of point clouds are one of the key technologies for model reconstruction.The PointNet network is a pioneering work for semantic segmentation directly using point clouds as input.In this paper,in view of the shortcomings of the original network's insufficient local feature extraction,the network structure is improved,and KNN(K-Nearest Neighbor)classification is added to the local feature extraction part.Algorithm,and make the underground pipe gallery data set to complete the point cloud semantic segmentation experiment,compare the experimental results with the original network and analyze the accuracy.Finally,the segmented point cloud data is used to quickly reconstruct the underground pipe gallery scene.The main work and research contents of this paper are as follows:(1)Point cloud data set production.There is currently no public underground pipe gallery data set.In order to carry out related research,this paper uses Cloud Compare software to perform point cloud labeling and creates a large three-dimensional point cloud data set according to the requirements of the segmentation module in the PointNet network.It includes 6 major areas and 8 main categories of point clouds in the urban underground pipe gallery scene.(2)An improved PointNet network is proposed,which is used to complete the semantic segmentation of point cloud scenes.The KNN neighborhood sampling algorithm is added to the local feature extraction process of the original PointNet network structure,and the point cloud features extracted by the KNN neighborhood sampling algorithm are input into the local features of the point cloud and merged with the global features to improve the overall segmentation of the network ability.The results show that the average intersection ratio(m Io U)of the proposed method and the overall segmentation accuracy(OA)of the original network are improved by2.4% and 2.7%.(3)Using the point cloud after semantic segmentation to quickly construct the underground pipe gallery model.Differential modeling is adopted for the different features of each object in the underground pipe gallery scene.The segmented point cloud data has the advantages of easy identification and rapid extraction of its features.This advantage can be used for simple and regular objects in the scene,such as walls.,Pipes and brackets to achieve rapid modeling.For the various complex small parts in the underground pipe gallery,the professional three-dimensional modeling software3 ds Max is used to construct the fine three-dimensional model,and finally the two types of models are merged and the model attribute information is given.