The 3D Real-Time Reconstruction From Point Cloud

3D reconstruction is a hot subject of image processing, belonging to multi-disciplinary subject, involved in computer graphics and image processing. It has been applied in many research fields, such as Manufacturing, Virtual Reality, Architecture, Medicine, Culture Relic protection, Film-acrobatic making and so on.This thesis surveys the current status of 3D reconstruction system, outlines its structure and module. Several key techniques existing are studied in this thesis, including the establishment of 3D multi-resolution triangular meshes from point cloud and the registration of multi-view range images. Finally we obtained a complete 3D geometric surface description of an object.The protracting of polygonal meshes is a standard method at present, which connects the points on the surface of 3D model into polygonal meshes. It can describe complicated surface and great applicability. An important problem in mesh reconstruction is how to obtain a reconstructed mesh with the same topology as the surface of the sampled object. In this dissertation, we discovered an intrinsic property of point cloud, and designed a reconstruction algorithm, based on analogical surface. The reconstructed mesh, in fact, is the analogical triangulated division restricted on the surface of the sampled object. Its topology is very close to that of the surface of the sampled object, and the mesh surface with the same topology as the surface of the sampled object can be obtained just with a simple post-process.The Iterative Closest Point (ICP) algorithm is a widely used method for aligning three-dimensional point sets. The quality of alignment obtained by this algorithm depends heavily on choosing good pairs of corresponding points in the two meshes, and error metric. Choosing good pairs of corresponding points is most important to improve the speed and convergence of the algorithm; and choosing the best error metric can enhance the precision of the final model pose. In this paper, the sampling stability and error metric have been analyzed and a solution has been proposed.