| Reconstruction of dynamic 3D models is one of the newest and hottest topics in current researches in computer graphics. As a method for 3D shape acquisition, the reconstruction of dynamic models concept do not require the scanned model to be hold still. Through capturing the motion of dynamic object real-time by depth scan device, its aim is to recover the continuous motion of object or people's expression, improving the ability of computer for exhibiting the real world.Non-rigid registration is the key work in dynamic 3D models reconstruction, also greatly affect the efficiency of process and details of result. Through pair-wise matching, we can accumulate the sampled point information of surface. Considering the most registration algorithm only suitable for small-difference deformed models, we present an efficient and robust pair-wise non-rigid registration algorithm for large-scale deformable models of dynamic objects. First, we use multiscale slippage analysis to detect salient features on source and target models, and use gauss curvature to match these feature points. After that, we apply an efficient pruning mechanism based on geodesic consistency to get high confidence sparse correspondence set. Then our approach performs the non-rigid registration by a global deformable optimization, guiding by these sparse correspondence and also dense correspondence generated by nearest close point method. The non-linear optimization maximizes the region of matching while minimizing the optimized energy function.Consequently, we combine the advantages of explicit correspondence computation and global deformable optimization techniques, and it could effectively handle large-scale deformation, match small details and fill holes. The registration results for real-scanned and synthetic data sets demonstrate that our approach is robust, and is applicable to complex scenes that could not be handled by algorithms assuming that inner-frame deformations are small. |
PDF Full Text Request