Building Extraction And 3D Reconstruction Of Multi-scan Terrestrial Laser Scanning Data

As one of the most important data sources,terrestrial laser scanning(TLS)can acquire rich 3D information on the target surface actively,quickly,accurately and without contact.The data obtained by terrestrial laser scanning are called point clouds.The original point clouds in the scanning scene commonly contain various types of objects in the scene,and have the characteristics of large amount of data,disorder and so on.Additionally,several scans are commonly needed to obtain the complete target surface because of the obstacle and occlusion in the scene.Thus,many steps are needed to convert the original point cloud to 3D geometric frame.Aiming at the 3D reconstruction,this study focuses on the point cloud index,registration,classification,building extraction,segmentation and 3D reconstruction during the the point cloud processing.Detailed research contents are as follows:(1)Point cloud index.This paper introduces and compares commonly used regular grid,octree and KD tree index in point cloud organization.A linear KD tree is proposed to reduce the memory requriment during KD tree building and neighboring search.The test shows the proposed linear KD tree can reduce the memory requirement and keeps a similar searching efficiency as the commonly used pointer-based KD tree.(2)Registration.The commonly used registration methods are analysed and compared in this study.Focusing on the problems of feature-based registration methods when dealing with the scene consisting of many repetitive and symmetric structures,an automatic registration method based on combination of smartphone and TLS scanner is proposed.The relative relationship between the adjacent scanner positions are measured by the smartphone and the distribution consistency between neighboring scans is measured by information entropy.The transformation parameters are found by searching for the best distribution based on the information entropy.Experiments on several datasets show that,the proposed method in this paper obtains a higher success rate and efficiency compared with the commonly used feature-based method,and is more suitable for the registration of urban building point clouds composed of many repetive and symmetric structures.(3)Building extraction.This paper analyzes and compares the methods of building extraction based on single point classification and point cluster classification.In view of the problems existing in the two methods mentioned above,a comprehensive building extraction scheme combining single point and point cluster is proposed.A method of point cloud density estimation based on scanning distance and angular resolution is proposed.In the classification,the density adaptive feature extraction method is proposed based on the density estimation method,aiming at the problem that the point cloud density changes across multiple orders of magnitude in the ground laser point clouds as the point cloud classification.Comparison experiments on building extraction on measured data show that the density adaptive feature proposed in this paper has a higher overall recall and accuracy in the point cloud classification,and has a better scanning distance applicability for building categories,and a better classification result is obtained in the extraction of long distance and low density buildings,so it is built in a large area of urban area.Building 3D laser scanning has better applicability.The improvement of density adaptive method also provides a new idea for improving the accuracy of building extraction from large scale terrestrial laser scanning data.(4)Segmentation and 3D reconstruction.An improved regional growth method is proposed to avoid the instability of the segmentation results caused by the selection error of the seed points and the uncertainty of the growth criteria.The method is based on the dimensionality elimination vector uncertain region,only in the region of the normal vector stable region,and then the removed area points are allocated to the nearest plane slices,and then the plane segmentation is realized.On the basis of the segmentation of the building point cloud,the contour extraction and semantic recognition of the different plane slices are carried out.The facade slices,based on the facade grid image,are extracted from the windows and other details,and the profile points are adjusted by the topological relations between the surfaces,the geometric frame of the building is constructed,and the geometric reconstruction of the building point clouds is realized.