Retrieval-based Multi-view Point Cloud Registration Study Method

Multi-view point cloud registration is to estimate the absolute poses of each point cloud for a disordered set of point clouds.Existing multi-view point cloud registration methods usually consist of two parts:pairwise registration and global consistency optimization.In this case,pairwise registration is to estimate the relative pose between two point clouds.Global consistency optimization uses these relative poses to estimate the absolute poses of each point cloud.Multi-view point cloud registration has been a challenging problem.The input multi-view point clouds are captured from different views of the scene,some of which have no overlap with each other.If pairs of point clouds with no or too little overlap are registered,many incorrect relative poses are generated.If a large number of erroneous relative poses are introduced into the subsequent global consistency optimization stage,it will lead to unreliable optimization results and degrade the multi-view registration performance.To reduce the generation of erroneous relative poses,this paper introduces point cloud retrieval into multi-view point cloud registration and proposes a retrieval-based multi-view point cloud registration framework.The retrieval module can retrieve similar point clouds and effectively filter out those that have no overlap,thus reducing the erroneous relative poses and improving the robustness of multi-view registration.In addition,in the global consistency optimization stage,the absolute poses of point clouds are estimated by using pose graph optimization,and the pose graph is iteratively optimized by using hierarchical registration to further improve the performance of multi-view point cloud registration.This paper compares various methods on the publicly available dataset Scan Net.Compared with the existing state-of-the-art multi-view registration algorithms,the mean rotation error and mean translation error of the multi-view registration framework proposed in this paper are reduced by 24.6~o and 0.76 m,respectively,and the CDF value are improved by 33.8%for the 10~o rotation threshold and 56.1%for the 0.25m translation threshold.The experimental results of multi-view registration show that the performance improvement of the proposed framework in this paper is very obvious compared with other methods.