Researches On 3D Reconstruction With High Precision

Our research is focused on the problems of 3D reconstruction from image sequences and some critical issues related to this subject, such as projective reconstruction and camera self-calibration. Traditional approaches are based on a preliminary calibration of the camera setup. This, however, is not always possible or practical. The goal of this work is to investigate the theoretical and practical feasibility in Euclidean reconstruction from image sequences without any prior knowledge either about the parameters of the cameras, or about their motions. Meanwhile, several methods, such as how to improve the accuracy and the speed of reconstructed 3D object, are considered in this thesis. Main contributions are as follows:1. An iterative factorization method based on 1D subspace for projective structure and motion is presented. It relies on the fact that the basis of projective subspace can consist of the three rows in the matrix comprising the first image points and one row vector that is orthogonal the former. The experiments with both simulate and real data show that the method is efficient for the recovery of projective structure and motion.2. A method for projective reconstruction from uncalibrated image sequence with occlusions is proposed. The re-projection points replace all the occlusions and projective reconstruction is obtained from all the points alternately, and the real positions of occlusions and the accurate projective reconstruction are finally obtained. Because the occlusions are obtained from all the image points, the method is precise for the recovery of projective structure and motion.3. To minimize the mean-squared algebraic distance and the geometric distance, we propose a linear resection-intersection bundle adjustment method for photogrammetric bundle adjustment. We adjust the parameters in so small range that the values of projection matrices and the space points can be substituted by their values in the last iteration. The theory and experiments show the convergence speed of the method is fast.4. A camera self-calibration method based on the linear iteration is presented. Based on the constraint equations of the absolute quadric and the hypotheses that the skew factor is zero and principle point locates the origin, the initially camera intrinsic parameters are attained. By the coordinate transformation, the image sequence accords with the one that the assumed camera captures. After several iterations, thereal intrinsic parameters are obtained. The experiments show that the self-calibration method is efficient, robust and has good property of convergence.5. The elements of the dual absolute quadric have so large differences in magnitude that solutions are extremely sensitive to noise. To solve the problem, we proposed the method that all the elements of the dual absolute quadric are transformed into the same magnitude based on an evaluative intrinsic parameter. The self-calibration method can lead to an enormous improvement on the stability and robustness of the results without increasing computation.6. We address an augmented reality system that operates in the environment that the translation of the camera is very small compared with the distant to the 3-D scene. This is a special case of the general problem, which significantly simplifies merging the real and the virtual, because a homography can describe the image motion between two frames of a video sequence. The theory and experiments with real video sequence show that the proposed AR system is very simple and efficient.