Research And Implementation Of Registration Algorithm For Large-scale Building Point Cloud

Point cloud registration is a basic and key technology in point cloud data processing and an important content in the fields of 3D reconstruction and 3D recognition.In the process of obtaining the building point cloud with 3D laser scanner,due to the characteristics of laser propagation along a straight line and the shielding between objects,in order to obtain the complete point cloud model of the building,it is necessary to set up multiple stations around the building to scan at different locations.The task of point cloud registration is to register independent point clouds in multiple coordinate systems to a unified coordinate system through rotation and translation operations,so as to obtain complete point cloud model.This research comes from the actual needs of enterprise and has important research significance.Aiming at the registration problem of large-scale building point cloud,this paper presents a registration framework of large-scale building point cloud,which includes three stages: coarse registration,fine registration and multi-view registration.Coarse registration is used for firststep registration between two pieces of point clouds,while fine registration is used for optimizing the result of coarse registration by iterative steps.Multi-view registration is based on coarse registration and fine registration and is used for registration among multiple pieces of point clouds.The main research contents of this paper are as follows:(1)In terms of coarse registration,point cloud preprocessing,feature detection,feature description,feature matching,mismatch filtering,transformation matrix calculation and other implementation processes are adopted.Aiming at the problem of poor filtering effect of RANSAC algorithm in the process of mismatch filtering,a mismatch filtering algorithm based on geometric compatibility was proposed to improve the effect of mismatch filtering.(2)In terms of fine registration,an improved ICP algorithm was proposed to improve the computational efficiency of ICP algorithm,aiming at the problem of poor real-time performance of ICP algorithm in the registration process of large-scale building point cloud.(3)In terms of multi-view registration,in view of the topological relationship formed by the adjacency relationship of the point clouds at the scanning positions,a multi-view registration algorithm based on topological relationship was proposed,and the improved shape growing algorithm,graph optimization algorithm and divide-and-conquer algorithm were adopted to solve the registration order and error accumulation problems of the point clouds with tree topological relationship,ring topological relationship and complex topological relationship respectively.This paper uses large-scale building point cloud data to verify the above algorithms.The experimental results show that the algorithms at each stage of the point cloud registration framework proposed in this paper are effective and have practical application value as they meet the actual needs of enterprise.