Coarse Registration Algorithm Of Point Cloud Based On Ridge-Valley Feature And Deep-Learning Describer

With the wide application of 3D reconstruction technology in the fields of reverse engineering,biomedicine and virtual reality,the digital processing of 3D objects in the reconstruction process has become a hot research topic.In the process of 3D object digitization,point cloud registration is very important,which can generally be divided into two steps: coarse registration and fine registration.Coarse registration uses algorithm to improve the overlap rate between two point clouds with large separation,providing a basis for the success of subsequent fine registration.Affected by the complex characteristics and noise inside the point cloud,the registration effect and robustness of the traditional algorithm have a large room for improvement.In recent years,the widespread popularity of deep learning has introduced new ideas for point cloud registration.This paper studies the point cloud registration problem by combining ridge and valley features and deep-learning describer,and improves the current feature description,feature filtering and feature registration methods,as follows:(1)A feature descriptor construction method for ridge-valley feature lines is proposed: for the problem of insufficient discrimination and robustness of feature descriptors,the algorithm uses ridgevalley feature lines with higher repeatability and feature information to construct feature descriptors.A set of descriptors is obtained by constructing FPFH descriptors for each point on the feature line,and the descriptor set is max-pooled to construct descriptors for the entire feature line.Experiments show that compared with the registration algorithm based on feature points to construct feature descriptors,the accuracy of the algorithm in this paper is higher.(2)A point cloud coarse registration method based on Spin Image descriptors is proposed: for the problem of mismatching in the Super4 PCS algorithm,the algorithm uses the Super4 PCS point cloud registration strategy based on Spin Image descriptor filtering.By calculating the Spin Image descriptor for the coplanar four-point base,then comparing the differences between the descriptors,setting a threshold to filter out the wrong matching pairs,and improving the registration accuracy.(3)A point cloud coarse registration method with single match point and deep-learning describer is proposed:(1)In view of the problem of poor repeatability of points to be matched due to random selection of points in the 3DMatch algorithm,the ridge-valley feature sampling strategy is used instead of random sampling selection.The points to be matched improve the repeatability of the key points and ensure the registration accuracy of the point cloud;(2)According to the characteristic that the feature points of the ridge-valley have both feature orientation and normal information,a local coordinate system is constructed for each feature point,which makes the point cloud registration can be achieved by mapping a local coordinate system between a set of matching keypoints;(3)An error de-filtering method based on a local-global ICP strategy is proposed,which improves the success rate of registration.