Research On Registration Technology Of 3D Point Cloud

With the rapid development of industrial technology and economy,reverse engineering has been widely applied in the fields of program evaluation of product,automatic manufacturing,management and maintenance.In reverse engineering,in order to get a complete 3D point cloud from a target object,3D point clouds collected from different views need be merged into a complete 3D point cloud by rotating and translating according to a transformation.The process is called as registration technology of 3D point cloud,which is not only the core technology in data processing of 3D point cloud and model reconstruction,but also the key technology in reverse engineering.The registration of 3D point cloud has been studied in this paper with feature point detection,feature description,feature matching,and elimination of mismatched pairs.The main contents include as follows based on the idea of "first coarse and last refine".First,the methods of feature point detection and description of 3D point cloud are focused on.A method of feature point detection combining surface variation index and eigenvalue variation index is proposed.The feature points detected by proposed method are not only invariant to rotation and translation,but also more repeatable than those detected by LSP,ISS and KPQ.A multi-scale covariance matrix descriptor used to describe feature point is proposed.By this descriptor,the multi-scale neighborhood information of a feature point can be accurately described in a low dimension,and its performance outperforms FPH,USC and Spin Image,etc.Second,in the research on feature matching of feature descriptor,a method of feature matching based on bidirectional nearest neighbor distance ratio is proposed.According to the proposed matching method and measurement method of log-eigenvalue,a multi-scale covariance matrix descriptor can be matched accurately with another one,and its matching performance is better than that based on nearest neighbor and that based on nearest neighbor distance ratio.Combined with feature point detection,feature description and feature matching,3D point cloud matching is performed for three different types of data.The results show that 3D point cloud matching in this paper can achieve better performance in the matching between a complete 3D point cloud and a complete 3D point cloud with noise,and between a single complete 3D point cloud and multiple complete 3D point clouds.However,in the matching between 3D partial point cloud and 3D partial point cloud,matching effect is not good.Third,according to analyze deeply the reason of mismatched pairs of feature point,the idea of eliminating mismatched pairs on the stage of feature point detection and feature matching is determined.Based on the characteristics that most of the neighbor points of a feature point are distributed on one side,a method of edge point detection is proposed.By using this method,a feature point locating on the edge of the 3D point cloud will be eliminated to decrease the number of mismatched pairs resulting from the edge feature points.For the matched pairs obtained from feature matching,a method of eliminating mismatched pairs based on k-means and splitting method is proposed.Based on these two methods at two different stages,an improved matching method between 3D point clouds based on feature is proposed,which can further effectively improve the matching efficiency and reduce the matching error.Finally,an automatic registration of 3D point cloud is focused on,including an automatic pairwise registration and multi-view registration.In the automatic pairwise registration,a coarse registration based on feature point detection,feature description and feature matching,elimination of mismatched pairs is stuied in depth,and a refine registration similar to ICP algorithm is studied deeply.In the automatic multi-view registration,a coarse registration based on shape growth and update,and a refine registration based on minimizing average distance of feature pairs between matched 3D point clouds are studied deeply.The experimental results show that the proposed automatic registration of 3D point cloud can run well over multi-view 3D point clouds,and its registration effect is better than that based on paired matching.