Abstract
3D modeling from images is an interesting and challenging problem in computer vision. Structure from motion (SfM) is a method that reconstruct 3D model from images. In this thesis, we propose a robust long-term SfM algorithm to reconstruct 3D models. This method overcomes the problems due to tracking errors and occlusion. For the problem of occlusions, especially occurs in the long-term sequence, we propose a new scheme for dealing with the missing data. Base on the idea of dividing image sequence into overlapped sub-sequences and then propagating points from the visible ones to the occluded ones. Those putative points are treated as the actual image points with some level of Gaussian Noise. Without any registration methods, this approach is straight-forward and works well in practice. To achieve robustness against outliers, we propose a robust SfM algorithm by applying the adaptive RANSAC technique on the projective factorization method. With slightly modifying the adaptive RANSAC algorithm, we obtain more inliers in one sample, thus reducing the maximal sampling times. Furthermore, to minimize the re-projection errors of the visible points considered as inliers, we propose a refinement algorithm to refine the model. Experimental results on both synthetic and real data show the robustness of the proposed algorithm.