Recovering 3D shape from 2D images, or 3D reconstruction, is not only the main task of human vision, but also a key research area and final aim in computer vision. Camera calibration is a key step in the process of 3D reconstruction. Any algorithm of camera self-calibration is used to compute the intrinsic parameters of camera just from feature-point set of views, which makes self-calibration work on-line, while traditional calibration techniques are off-line. Therefore camera self-calibration has been quite valuable and promising both in theory and application, and this technique has been arisen as one of the most important specializations in computer vision since late 1990s.In this paper, we focus our attention on the research and exploration of camera self-calibration and 3D reconstruction. We start our work from model-based calibrated single view, and after the correspondence relationship between the null space fell by the depth vector of the target imaged in the single-view and the null space of the model shape matrix of the target is built, a new algorithm to linearly and exactly reconstruct the 3D pose of the imaged target from single-view Is presented. The algorithm needs at least six points to get the solution. The theoretical analysis and a great deal of experiments have demonstrated that the suggested algorithm is fast, efficient, effective and rather robust to noise. Then we step further and generalize '6-point algorithm' to the un-calibrated single-view, where 'Model-based linear self-calibration and 3D reconstruction from single-view' algorithm is presented. This technique has inherited the strongpoints which are possessed by '6-point algorithm', and it also succeeds in accurate calibration. For the sake of gaining higher accuracy in calibration, we forward our attention to multi-view and generalize the algorithm presented above to a multi-view one, where 'Model-based linear self-calibration and 3D reconstruction from multi-view' algorithm is presented. Experiments have demonstrated that the suggested algorithm has gained much higher accuracy in calibration. Besides, we also present two ways to realize this technique with contrasted experiment results attached.As we know, successful self-calibration techniques from multi-view are... |