The study on 3D modeling based on 2D video sequence is of great significance in computer vision and biomedical image processing. In order to satisfy the pressing needs of gastrointestinal tract diseases diagnosing and treating in clinic medicine, the gastrointestinal tract 3D modeling system is studied systematically in this dissertation. The focus is put on some of the fundamental problems of 3D modeling approaches and systems, including scene manifold modeling, low-level processing algorithms, 2D and 3D image processing and physical implementation of new type wireless capsule endoscopy (WCE). The main works are summarized as following: (1) The architecture of the gastrointestinal tract 3D modeling system is established. (2) The manifold modeling approach and its application in 3D modeling are proposed and analyzed in detail for the first time. Actually, the concept of manifold in mathematics can be reinterpreted with camera projection in computer vision, and both the image mosaicing algorithm and local coordinate registration methods can be described in a unified way by manifold modeling. (3) The scaled total least squares (STLS) algorithm for linear model parameters estimation is implemented and its topological interpretation and relation with Levenberg-Marquardt algorithm are also presented. Furthermore, a novel estimation algorithm for intra-frame homography is put forward by combing the hierarchical motion model and optical flow equations, which works well when all the feature-based homography estimation methods fail. (4) The tubular scene manifold mosaicing (TSMM) algorithm and its revised version are proposed, which are applicable for the gastrointestinal tract scene and other scenarios in which the camera shots in an axial-way. The investigation shows that: the image acquisition procedure exerts a tremendous influence on the design strategies of image mosaicing; image strip selection and reshaping are two necessary steps in mosaicing algorithms; the orthogonality constraint and topological connectivity should be satisfied in image strip reshaping operation; and lastly, the complex geometry and Mobius mapping are valuable for the implementation of strip reshaping. (5) An iterative algorithm for adding views in metric 3D reconstruction is presented based on two-view 3D reconstruction. The key of the algorithm is the recursion formula for local coordinates alignment. The 2D+3D modeling method is proposed to combine the advantages of image mo...
修改评论