Research On Registration Technology For 3D Point Cloud

With the rapidly development of modern information technology and graphics,3D reconstruction technology gains more and more attention in reverse engineering,robot vision,medical imaging,3D printing,digital entertainment,and other fields.The goal of it was to build three dimensional digital model of the object by processing the surface point cloud obtained from scanning device.But only one scan cannot get all of the point cloud data due to the influence of scanning device,measurement conditions and the object itself.To get a full digital model,we need multiple scan to obtain point cloud data for subsequent processing.3D point cloud registration is looking for the mapping relationship between different point clouds and then unified them to the same coordinate.The key of it is calculating the transformation parameters: the rotation matrix and the translation vector.As an important link of 3D reconstruction,3D point cloud registration technology has always been a hot research topic.The paper makes a deep research on previous work and design two kinds of method for point cloud registration.First of all,the article made a brief introduction about the background and research situation at home and abroad of the subject research.Through analyzing the research situation,registration algorithm can be divided into three categories: the algorithm based on the geometric feature,the method based on statistics and the method of iterative least error.Meantime,the paper introduced the basic knowledge of point cloud registration to lay a foundation for subsequent algorithm design.Secondly,the article puts forward a registration method based on the improved Coherent Point Drift(CPD).CPD algorithm regards the registration of two point clouds as a probability density estimation problem,where one point set represents the Gaussian Mixture Model(GMM)centroids,and the other one represents the data points.It fits the GMM centroids to the data by maximizing the likelihood to align the two point cloud.But the algorithm has a large amount of calculation and the low registration efficiency when deal with a massive point cloud data.So the voxel grid method is used in this paper to reduce the size of point cloud data for calculating.This method gets good registration accuracy and improves the matching speed.Thirdly,a new registration method based on multi-scale axis angle feature was proposed.The innovation of method is mainly reflected in these aspects: key points were selected based on mean value of projection distance along normal;the angles deviation between axes ofdifferent local coordinate system and curvatures were made the key points feature descriptor;using the descriptor similarity constraint to determine initial correspondence,and optimizing the correspondence by using the random sample consistency algorithm and the clustering selection method;the rotation and translation matrix were estimated by SVD(singular value decomposition).Finally,in the VC++ environment,it designed experiments to validate the feasibility and effectiveness of two registration algorithm by a variety of point cloud data.The experimental results show that both methods can achieve the desired effect.The point cloud registration method based on multi-scale axis angle feature has low computing complexity,fast registration speed and high precision.The method has practical value in the actual engineering application with its anti-noise ability.