Researches On 3D Reconstruction From Uncalibrated Image Sequences

The Structure from Motion ( SFM ) problem is the central problem in computer vision community. The uncalibrated structure from motion problem ( USFM ) is the hot topic in the recent 10 years. Many theories, including multiple view geometry, protective reconstruction and camera self-calibration etc. have been proposed and well developed. In the applications, as the technologies of computer and internet developing, many applications require 3D graphic data of high quality. While previous CAD methods for 3D graphics are expensive in time and labor consumption, and the obtained 3D graphic models are often too simple for applications. One challenging problem is how to obtain complicated and realistic 3D graphic models as automatically as possible. This thesis is focused on this problem and improve some key algorithms in the USFM problem. The effectiveness of our algorithms are evaluated in the experiments with simulated data and real image sequences.1. In the image matching algorithm, a RANSAC homography method is proposed to improve the efficiency of coplanar feature matching. It is very suitable for the coplanar dominant case such as building image sequences. In the accurate estimation of epipolar geometry, a optimization algorithm for estimating Fundamental Matrix by integrating coplanar and non-coplanar correspondences. It makes the obtained fundamental matrices more accurate in the bad feature configuration in building sequences.2. An analysis on camera critical motion is given and a conclusion is proposed which clarifies that the critical motion is decided by the ambiguities of the absolute quadric. Comparing to the previous definition, our definition not only useful for the case of variable intrinsic parameters, but clearly show that critical motions are mainly caused by extrinsic parameters.3. A new theory and algorithm for camera auto-calibration is proposed by integrating some prior scene knowledge. The algorithm transform the line parallelism and plane orthogonality into some new constraints on auto-calibration equations, which makes the affine and Euclidean characteristics of the absolute quadric close to the practical one. This results the parallel and orthogonal structure of the reconstructed scene more accurate. Furthermore, the new constraints decrease the ambiguities in solving the absolute quadric, which improves the robustness of the previous algorithms to camera critical motions.4. A triangulation algorithm integrating the 3D depth information of the scene isproposed. It makes the detail characteristics of the reconstructed scene better. While the previous algorithms often produce the reconstructed 3D mesh with some large holes.5. By analyzing on the camera single axis rotations, we concluded that there are ambiguities in the 3D reconstruction. A corresponding approach is proposed to remove the ambiguities. This kind of method for 3D reconstruction requires a simple and feasible system, which is suitable for modeling small objects.