The study on robust 3D reconstruction techniques is one of the main research directions in 3D computer vision, and multi-view based reconstruction is considered as the most effective way to increase the robustness aspect. The thesis is focused on robust trinocular image rectification, stratified reconstruction and calibration of a structured light based 3D reconstruction system, the main work is summarized as follows: 1. A robust trinocular rectification method is proposed. At first, the necessary and sufficient condition for trinocular rectification is obtained, and a linear rectification algorithm is proposed. Then a RANSAC based algorithm is introduced, which is shown by experiments to be sufficiently robust. In addition, a geometric interpretation of the 6 free parameters involved is also provided, by which the induced projective distortion is significantly reduced. 2. A new stratified reconstruction algorithm based on affine point correspondences is proposed. The affine point correspondences here mean image correspondences which are projected by two sets of space points related by a general 3D affine transformation. The proposed algorithm consists of the following three main steps: First, the quasi-affine reconstruction is obtained from correspondences of image points; Second, the 3D projective transformation matrix between the two sets of quasi-affinely reconstructed points in the first step is computed. It is proved that among the 4 eigenvectors of this transformation matrix, one eigenvector must correspond to the plane at infinity. By determining this eigenvector, the quasi-affine reconstruction is upgraded to a true affine reconstruction; Third, the camera intrinsic parameters are calibrated from the infinite homography obtained in the second step, and a metric reconstruction is obtained. Moreover, we also prove that if the camera undergoes a planar motion, then the affine reconstruction is impossible. In other words, the planar motion is a kind of degenerate camera motion in this case. 3. Three different system calibration methods are compared for a structured light based 3D reconstruction system. Although the fixed-point based method and the fixed-line based method are elegant from mathematical point of view, the traditional coordinate transformation based method is found be experiments to consistently outperform the other two in term of robustness, hence is highly recommended in real applications.
修改评论