Research On 3D Point Cloud Registration AlgorithmBased On Geometric Features

With the development and popularization of 3D point cloud technology,3D point cloud image processing technology plays an irreplaceable role in 3D reconstruction,medical imaging,virtual reality,driverless and other fields.Because the 3D scanning device can not obtain all the point cloud images of the object at one time,it is necessary to register the point cloud obtained in different directions to the common coordinate system to obtain all the point cloud data,which is called point cloud registration.Three-dimensional point cloud registration is the most basic and important part in 3D point cloud image processing,which consists of coarse registration and fine matching.In this paper,coarse registration and fine registration algorithms are studied respectively,and the shortcomings of the present registration algorithms are analyzed,and the algorithm is optimized from two aspects:registration accuracy and registration time.The main work of this paper is as follows:Firstly,the research significance of 3D point cloud registration is introduced,the present research situation at home and abroad is summarized from two aspects of coarse registration and fine registration,the shortcomings of the current registration algorithm are summarized,and the related processes and methods of point cloud registration are expounded.Secondly,a rough registration algorithm based on FPFH(Fast Point Feature Histogram)feature is studied in depth.A rough registration algorithm based on curvature and FPFH feature is proposed to optimize the algorithm from two angles of registration error and time.First,the key point set of point cloud is determined by the curvature feature,which reduces the data redundancy and computation cost;second,the FPFH features of key points are calculated and the corresponding point sets are determined;the RANSAC(Random Sample Consensus)algorithm is introduced instead of the SAC-IA(Sample Consensus Initial Aligment)algorithm to optimize the corresponding point set,which is helpful to obtain the accurate transformation matrix and improve the registration accuracy.The experiments on Stanford datasets and datas collected by kinect show that the algorithm can effectively reduce registration error and speed up registration.Thirdly,in order to solve the problems of slow searching,many wrong corresponding points and poor registration efficiency of traditional ICP(Iterative Closest Point)algorithm,this paper proposes an ICP algorithm based on geometric feature.First,the point cloud is simplified by using normal vector features,which reduces the calculation of registration process;second,the nearest neighbor search is accelerated by using kd-tree,which improves the efficiency of point pair search;finally,the corresponding point set is determined by introducing FPFH feature,which improves the proportion of the correct point pair.The experiments on Stanford datasets and datas collected by kinect demonstrate that the proposed algorithm significantly improves registration performance compared with conventional ICP algorithm.Finally,the work done in this paper is summarized,and the future research work is prospected by analyzing the shortcomings of this work.