Distance Field Based Non-rigid Registration And Modelig

3D modeling is a fundamental research direction in computer graphics. The goal of3D modeling is to convert a virtual or real object to a geometry represen-tation.3D models is the basis of further processing and maniputation. Therefore,3D modeling is widely used in rendering,computer animation, human computer interaction, computer aided design, etc. The common3D modeling method can be devided into interactive modeling and modeling based on3d scanned data. With the rise of3D scanner, the amount of3d scanned data is increasing. Modeling based on3D scanned data is becoming a hot research area.Registration is a fundamental problem in modeling based on3D scanned data. Registarion of3d data is to align several3d datasets with different transform to a common coordinate. Registration could be devided into rigid registration and non-rigid registration. Rigid registration is to recover the rigid transform between the shapes. Non-rigid registration is to optimize the non-rigid transform between the shapes. With the emergence of tiny, portable3D scanner, the amount of temporal shape sequence is increasing, that leads to oppotunities for non-rigid registration.In this paper, we propose a non-rigid registration based on distance field. With the locally rigid, globally non-rigid assumption, we present a volumetric framework for non-rigid registration.First, we investigate the existed frameworks for non-rigid registration. For shape representation, we studied the implicit surface and explicit surface. For local geometry matching, we analyzed the performance of nearest point based method, feature based method, as well as patch based method. For deformation model, we investigate the linear and non-linear models, surface and space based models. For gloabal deformation optimization, we describe the locally rigid, globally non-rigid framework.Second, we implemented and improved a GPU based signed distance field generation algorithm. We analyze the algorithm from the definition of signed distance function, implemented the algorithm for GPU based unsigned distance field computation. We also proposed an method for computing angle weighted pseudo normal using GPU. Using this algorithm, we could generate the sign of distance field.Then, we present a non-rigid registration method based on signed distance field, including cell based locally rigid registration and volume based deformation model. For the locally rigid registration part, we employ a3D Lucas-Kanade al-gorithm, optimize the rigid transformation by minimizing the difference of source and target distance field. For the globally non-rigid deformation model, we min-imize the difference between the adjcent cells and achieve a natural global de-formation. For the global non-rigid optimization, we propose a global energy function with confidence for each cell. It could reduce the influence of the wrong correspondence.Finally, we test our algorithm on several datasets, including synthetic data and real data. We proposed a volumetric method for accuracy measurement. And we compare our result with other method. Then we analyze the properties of our method and some future work.