Research On Automatic Registration Algorithm Of Three-Dimensional Point Cloud Models

The rapid development of the 3D laser scanning technology makes it possible to obtain the 3D point cloud model with large scale and high precision.Since the point cloud model is easy for storage,can draw the complex surfaces and 3D scenarios,and it does not contain complex topological connection information,it has become a mainstream representation in 3D model processing.To obtain a complete point cloud model of a real object,point cloud from different viewing angles in different coordinate systems must be registered.Therefore,registration results directly affect accuracy and efficiency in model reconstruction.This thesis analyzes and studies the initial registration and accurate alignment and proposes different improvement strategies aiming at accuracy and efficiency in automatic registration of the 3D point cloud model.This thesis mainly covers the following aspects:(1)To greatly improve the efficiency of point cloud registration,a RANSAC algorithm based on regional feature is proposed for initial registration.Firstly,the regional feature of a sampling point is defined based on the Delaunay neighborhood region and thus the feature point is extracted from the point cloud,and matching point pairs are selected based on initial features.Secondly,according to the determination of initial point pairs and the point pair filtering algorithm based on zero-mean normalized relationship,input parameters for registration algorithms are selected from feature points.Finally,the measurement standards for point cloud consistency are defined and the improved RANSAC algorithm is adopted for initial registration of point cloud.(2)To effectively improve the accuracy of point cloud registration,the ICP algorithm based on regional expansion is proposed for accurate alignment.Firstly,a set of feature points that correspond to each other from the point cloud to be registered are filtered based on the regional curvature similarity and distance restrictions among feature points.Secondly,a breadth-first search algorithm with pruning based on regional expansion is adopted to further search for matching point pairs,and redundant feature point pairs are excluded.Finally,the result of the initial registration algorithm is used as the initial condition for ICP iteration,and the quaternion algorithm is used to calculate the rigid transformation parameters between matching points.After multiple iterations,the optimal rigid transformation parameters are worked out for accurate alignment of the point clouds.(3)This thesis has implemented the algorithm on some point clouds with diverse features.The experimental results indicate that the automatic registration algorithm of the point cloud proposed in this thesis has a good effect,not only realizing accurate registration of point clouds,but also effectively avoiding invalid matching brought by similar feature points.This allows higher accuracy of the registration result.At the same time,the algorithm has been applied in the virtual restoration of Terracotta Army.