A Method Of Rigid Registration Based On Poisson Reconstruction Of Local Surface Sample

The registration of 3D point clouds is the hot issue in the fields of reverse engineering and computer vision etc.In order to improve the match accuracy of the corresponding point in the common areas between the floating point clouds and the fixed point clouds during the process of point clouds registration,this thesis proposes the matching rule by searching the nearest point of the local sample points to the Poisson surface as the corresponding point.The study uses the local sample data of the fixed point clouds to construct the Poisson surface and the nearest point of the sample point in floating point clouds to the Poisson surface as the corresponding point,which can extend the matching range of the corresponding points and change a single candidate corresponding point into infinite ones on a surface.In the test,the proposed method of registration can reach to the mainstream registration accuracy,speed up the convergence and improve the robustness of registration.The main contents and research results are as follows:(1)In order to improve the quality of reference surface—Poisson surface in the process of registration,a new method of Poisson reconstruction is presented by constructing the closed point set.By adopting the boundary identification method based on the normal projection of the triangle,the two-dimensional boundary feature and the three-dimensional boundary feature of the point set are acquired.It can turn sample data of original non-closed into closed point set by inserting discrete points,which can improve the quality and robustness of the Poisson reconstruction.(2)In order to effectively separate the mesh surface corresponding to the sample data,a new dynamic spatial index structure is proposed by combining KD tree and half-edge structure.Based on the openness of leaf nodes on the KD tree,the geometric information of the mesh surface can be stored in leaf nodes.The leaf nodes can be searched by the superior neighbor query performance of the KD tree.Topological neighborhood information of the target point is acquired rapidly by the ring sequence of half-edge structure and the dual side information.Based on the dynamic index,the mesh surface information corresponding to the sample data can be quickly separated,which ensures the topological integrity of the mesh surface and provides a more accurate reference information for registration.(3)Initial registration can be conducted iteratively by the interactive selection of feature points.The corresponding points of sample data from floating points and fixed points can be established by the match rule of point to Poisson surface.Apply the SVD to solve the parameter of the rigid transformation and define registration adjustment factor,which can adjust the range of candidate surface mesh nodes dynamically.The factor can make up the defect of low precision of manually selected feature points and improve the accuracy of initial registration.On the basis of initial registration,fine registration selects feature points and increase their number based on the adaptive strategy,which can guarantee the implementation of fine registration if the result of initial registration is not ideal.Therefore it can ensure the registration precision,improve the robustness of registration and reduce the number of convergence.(4)A new method of radial error estimation based on the distance of point to mesh surface is proposed.The fitting degree of floating point clouds and fixed point clouds can be improved by reducing the radial error.The adjusting factor can adjust the circumferential error caused by threedimensional measurement system to improve the accuracy of registration.The compensation error introduced can eliminate the error caused by Poisson surface reconstruction,making the error estimation much better in line with actual measurements and improving the accuracy of error estimation.