Research On Non-rigid Point Cloud Registration Based On Sparse Prior Correspondences

This thesis mainly studies the non-rigid registration.Existing methods are mainly based on the probability method and has made some progress.However,the registration performances can be unstable with a large deformation.To tackle this shortcoming with a non-rigid point cloud registration method based on sparse prior correspondences.First,robustly sparse correspondence between the source point cloud and the target point cloud are established through a global and local combined multi-level processing;then two constraints are constructed base on sparse prior correspondences,especially the local transformation constraint,which makes the local deformation of the point cloud as smooth as possible during the registration process.Specifically,this thesis mainly includes the following:First,the thesis reviews the background and significance of point cloud registration,the research status at home and abroad,and the relevant theoretical knowledge involved in the proposed method.Among them includes: the general process of point cloud correspondence estimation,point cloud registration process,probability density model,maximum expectation algorithm,point cloud super voxel segmentation algorithm,which lays the foundation for the introduction of our method.Secondly,a sparse correspondence estimation algorithm based on multi-level evaluation is proposed to get reliable prior correspondences automatically.The algorithm builds upon a global and local combined view,which gradually filters out incorrect matching pairs between the source point cloud and the target point cloud in a multi-level form,thereby obtaining relative robustly sparse prior matching pairs.First,the feature descriptors for all points from both the key point set and the target point cloud are computed and then the initial sparse correspondence according to the similarity of the feature descriptors are created;then,the source point cloud and the target point cloud are jointly segmented to get the regional correspondences between two point clouds.Two corresponding points not located in the corresponding local regions will be removed to keep global consistency.Then,those putative correspondences are further filtered by the local isometric and geometric structure.The combined strategy obtains the robust sparse prior matches.Experiments show that the method is reasonable and effective in obtaining sparse prior matches.Finally,a non-rigid registration algorithm based on sparse prior matches is proposed to tackle poor registrations due to large deformations.The algorithm is based on the sparse prior correspondences,and adds two constraints to the probability-based registration framework: one is a data constraint to minimize the distance between corresponding point pairs,the other being a local transformation constraint to keep the local deformation of the point cloud as smooth as possible.The experimental results show that the proposed non-rigid point cloud registration algorithm based on sparse prior correspondences is effective for large deformations.