Research On Point Cloud Registration Based On Optimal Transport Theory

With the rapid development of computer graphics,point cloud registration plays an important role in the field of computer graphics and has received more and more attention.Point cloud registration aims to process point cloud data which are obtained by multiple scans,exist in different coordinate system,and have no corresponding relationship.Because of the impact of light and the environment,there are noise and missing points in these data.Point cloud registration is design for overcoming the problems in the data and showing more realistic object model information through the registration.This thesis found the key points and difficulties in the research of point cloud registration algorithm through summarizing and analyzing the existing point cloud registration algorithms.Regarding the above problems,point cloud registration algorithm based on optimal transport and the concept of point cloud mass-reallocation is proposed.The mainly research content is as followings:1.Point cloud registration algorithm based on the theory of Optimal Mass Transport(OMT)is proposed,namely OMT-Reg(Optimal Mass Transport Registration)algorithm.The algorithm uses the cost of OMT to represent the distance between the point clouds and applies the optimal transport distance to measure the similarity of the two point cloud distributions.The proposed algorithm transformed the complex point cloud registration problem into a simple optimization problem.Through the establishment of the objective function,the OMT-Reg algorithm was used to optimize the distance between the point clouds.And finally,the objective function was used to achieve the registration between the point clouds.2.The concept of the point cloud mass-reallocation is proposed.The difficulty of the point cloud registration is that when there are many noise and missing points the registration results is inaccurate.Aiming at this issue,the mass of the point of the point cloud is adjusted based on the OMT-Reg algorithm.The concept of mass-reallocation proposed in this thesis obtains more accurate registration results through reducing the mass of the noise and missing points and reducing the influence of noise and missing points about the point cloud registration.The improved algorithm has high stability and accuracy.Experimental results show the registration algorithm which applies the mass-reallocation has good robustness to the existence of noise and missing points.