Research On Three-dimensional Point Cloud Coarse Registration Algorithm Based On Corresponding Point Reduction

Three-dimensional point cloud reconstruction is one of the key technologies of artificial intelligence and virtual reality.With the optimization of data collection equipment and the continuous improvement of computing power,it has drawn more attention from research and practical applications.3D point cloud registration is a key step in 3D point cloud reconstruction.Due to large amount of data to be matched in 3D point cloud registration,and the real-time performance requirements,high accuracy and high efficiency 3D point cloud registration is currently a main research direction in this community.This paper mainly focuses on the algorithm of key-points extraction and feature descriptors step in the coarse registration,and studies on registration accuracy and time.In this paper,based on the Point Cloud Library platform,the combinations of the SIFT and ISS classic algorithms for key-points extraction in the coarse registration with the classic PFH and FPFH algorithms for feature descriptors are studied.In terms of the relative repeatability rate of key-points and the ratio of the number of key points to the number of original point cloud points,the SIFT and ISS key-points extraction are compared.The experimental results of this paper show that the SIFT algorithm is superior to the ISS algorithm in terms of relative repeatability;the algorithms of different key-points extraction and feature descriptors are compared.The combined point cloud coarse registration result shows that ISS combined with FPFH point cloud registration algorithm has less calculation time and better overall performance.As the implementation of 3D point cloud coarse registration involve too many data points in each iteration,the corresponding point pairs are not refined,resulting in a relative long registration time,a simplified point cloud coarse registration algorithm for corresponding points is proposed.By optimizing the calculation of scene point cloud feature descriptors,reducing the selection of corresponding point pairs,and aiming at the feature that this algorithm only uses the central area to calculate corresponding point pairs,a corresponding point selection optimization algorithm is proposed to ensure that the corresponding points mostly do not deviate from the template point in other areas of the cloud,leading to better accuracy of corresponding point pair selection.The experimental results show that the coarse registration time of the algorithm proposed in this paper is reduced by 11.78%.Furthermore,although the coarse registration accuracy is sacrificed to some extent,the overall registration accuracy of the point cloud is little affected,and the overall performance of the point cloud registration is acceptable.Our algorithm might provide a possible solution of point cloud registration for real-time desirable applications.