Research On Point Cloud Registration Technology With Small Overlap

As people's perception of the world is not limited to two-dimensional information,the three-dimensional information of objects plays an increasingly important role in many fields such as three-dimensional reconstruction and machine vision.Due to the limitations of the shape,the size of the object and the range of viewing angles of the three-dimensional scanning device,scanning once cannot obtain the complete three-dimensional model information.To solve this problem,it is necessary to register the point cloud data obtained from different perspectives.That is,the correspondences between two point clouds at different angles of view should be detected in order to calculate the rotation and transformation matrix and unify the point cloud through them.This paper divides the point cloud registration method into two aspects after a great deal of research and summary of previous work.For the registration of the lower overlap rate point cloud,the article begins with two aspects: registration algorithm based on local high-dimensional features and Super 4PCS global registration method based on key points selection.Researches and improvements are made and the feasibility and effectiveness of the registration method are verified through experiments.Firstly the background,significance,application prospects of the point cloud registration technology and the recent research status at home and abroad are introduced.This article clarifies the concept of point cloud registration,describes the point cloud acquisition method,and relevant theoretical knowledge in the point cloud registration process,which paves the way for the follow-up research.Secondly,a key point extraction algorithm,the difference of multi-scale weighted normal projection,is designed after analyzing the algorithm based on normal vectors of points.Simulate different scale space around the query point and calculate the weighted projection mean difference between the point and the query point in the normal direction.The difference is taken as the basis for the selection of key points,which can effectively extract the points where the surface of the object changes significantly,and provide a higher-quality set of key points for subsequent processing.And SHOT feature is used to describe the key point set,the correspondences between the features are found to calculate the rigid transformation matrix for coarse registration,and finally ICP is used for accurate registration.Then,after analyzing and summarizing the Super 4PCS registration algorithm,an improved Super 4PCS registration algorithm based on key points is designed.Firstly,the original point cloud data is down-sampled through the voxel grid,and the key points are extracted through ISS algorithm.This can reduce the search range of the consistent four-point set,the time complexity of the algorithm and effectively suppress the noise.For the case that the point clouds with low overlap may result in poor registration,an overlap region extraction method based on the recently consistent four-point set is designed.The overlap region is used for ICP registration and the resulting rigid transformation is used for the initial point cloud to complete the registration.Finally,the experiment is carried out in the VC++ environment and experiments are carried out on different models of different methods to verify the accuracy and effectiveness of the proposed algorithm.