Research On Point Cloud Registration Algorithm In 3D Reconstruction

3D reconstruction can easily capture the appearance and 3D shape of the object,and simulate 3D interaction and perception in digital space.The goal of point cloud registration is to align one point cloud with another to estimate the best relative transformation.It is widely used in dense scene reconstruction,SLAM,tracking,and medical imaging.Depending on the registered objects,point cloud registration can be divided into rigid registration and non-rigid registration in 3D reconstruction.However,the existing point cloud registration algorithms often fail in complex scenes and cannot be well applied to 3D Reconstruction.To improve the accuracy and robustness of existing registration algorithms,and achieve better 3D reconstruction,this paper comprehensively analyzes three aspects of transformation model selection,corresponding constraints,and optimization methods,and conducts research on the point cloud registration algorithm in 3D reconstruction.(1)For rigid 3D point cloud,we proposed a probability re-weighted registration algorithm based on the Gaussian mixture model(GMM),aiming to solve the problems of massive outliers and missing correspondences in partially overlapping point cloud.Firstly,the correspondences between the target and source point clouds are established by the GMM and uniform distribution.We show that the missing correspondences in the target point cloud can be handled by re-weighting the mixing proportion of GMM through a prior probability reweighting strategy.Secondly,a posteriori probability inference strategy is used to infer the outliers and their proportion in the source point cloud,where the potential outliers are removed when solving the GMM parameters.Finally,the objective function in the form of point-to-plane distance is introduced by calculating the normal direction in the vicinity of the weight-averaged target point,and the expectation maximization algorithm is used to solve GMM parameters to finely register point clouds of large planar structures.Comparative experiments are conducted on Stanford 3D Scanning data and real 3D scene data,the evaluation results demonstrate that our algorithm is efficient for missing correspondences and outliers in 3D point cloud registration and improve the registration accuracy.(2)For non-rigid 3D point cloud,we proposed a non-rigid registration algorithm based on a curvature adaptive deformation graph and multiple geometric pruning,which is used to provide reliable correspondences and initialization poses for non-rigid registration,and solve the problems of poor performance of existing methods in the face of large deformations and missing correspondences.Firstly,by sampling the source point cloud based on the Gaussian curvature and local geodesic distance,a node graph reflecting the deformation of the source surface is constructed adaptively.The nodes can be evenly distributed on the source surface by sampling the key points of the deformation and controlling the sampling density.Secondly,the point cloud SHOT feature and curvature are used to find the initial correspondences,while the reliable correspondences are obtained according to the diffusion pruning technology.Finally,the correspondences are re-searched during the registration optimization,and the mismatches are removed to constrain the deformation domain according to the distance and normal pruning.Comparative experiments are conducted on the MPI-Faust and Human-motion datasets,the evaluation results show that the proposed algorithm can obtain better initialization poses and remove missing correspondences.While the average registration error is reduced by 1 to 5 times,the running speed of non-rigid registration is significantly improved.