As the developing of computer and internet technologies, there is more and more demand for 3D object or scene models in our life. Obtaining realistic 3D models from uncalibrated image sequences is a hot research aspect in computer graphics and computer vision fields since it is inexpensive and simpler.There are many key techniques in 3D reconstruction from uncalibrated image sequences, which are feature matching, fundamental matrix estimation, camera self-calibration, dense stereo matching and Euclidean reconstruction. Fundamental matrix and camera self-calibration are two key concepts, and model parameter estimation and implement of 3D reconstruction are main items as well. The thesis is focused on above associated issues.RANSAC algorithm is a classical method of robust model parameter estimation. A preview model parameters evaluation RANSAC (PERANSAC) algorithm is developed, which introduces a preview model parameters evaluation in the RANSAC algorithm, and results in better performance since the worse contaminated samples are discarded with guaranteeing the same confidence as RANSAC.After the non-algorithm factors of impacting fundamental matrix estimation are discussed and the fundamental matrix estimation with the uncertainty of the matches is investigated in detail, two new robust estimation algorithm for fundamental matrix are presented. One is based on PERANSAC, and the other is based on clustering, which utilizes improved SMEM algorithm to resolve the Gaussian Mixture Model that describes the residual distribution of matches. Both methods provide good performance, and the latter is better since it can give a better describing of the residual distribution of matches.In the study of camera self-calibration, Firstly, based on discussion of using and the selecting of scene constraints, two new robust vanishing point estimation algorithms are presented. One is based on PERANSAC, and the other is based on clustering. The latter can provide accuracy than the former. Secondly, a new method with scene constraint and the constraint of principal point for camera self-calibration based on the linear iteration is presented. It provides robustness in camera self-calibration.At the end of the thesis, a 3D reconstruction system is implemented, and the issues in the implement of 3D reconstruction system are analyzed. Some 3D scenes reconstructed by this system are shown.