Abstract
In this thesis, we propose a novel algorithm to simultaneously solve the stereo rectification and distortion correction problems in an integrated optimization framework. Stereo vision estimate depth information from a rectified image pair by finding point correspondences across images, i.e. stereo matching. Previous researches focused on developing accurate stereo matching methods. However, stereo rectification has been a crucial but overlooked topic to generate rectified images for stereo matching. The stereo rectification is especially important when the stereo images are acquired by different cameras with considerable lens distortions, which may lead to errors in image rectification and depth estimation. Considering different intrinsic and extrinsic parameters for stereo cameras, rectification homography transforms for both images are estimated based on the epipolar constraints. Lens distortion is also included into the generalized epipolar constraints by employing a division undistortion model and a lifting technique. By using the robust estimation formulation, we can solve the stereo rectification and distortion correction problems simultaneously in a unified framework. Experimental results on various synthetic and real images are shown to demonstrate that the proposed algorithm not only effectively reduces the rectification errors but also corrects the lens distortion in the stereo images. In addition, our experiments show that the proposed stereo rectification algorithm leads to more accurate disparity estimation results.