Based On Monocular Vision, Motion Estimation And Structure Reconstruction

Estimation of the camera motion and the structure of a scene from two-dimensional images taken by a camera is one of the core problems in the area of computer vision, which is called Structure from Motion (SfM). Reliable algorithms for motion and structure estimation are invaluable in visual robot navigation, cartography, aerial reconnaissance, target identification and virtual reality in computer graphics. Distinguished from stereo algorithms based on a traditional binocular stereo visual system or multi-view stereo visual system, the algorithm described in this thesis is based on monocular vision, the advantage of which is that it greatly decreases the equipment requirements and can be applied to an uncalibrated amateur handheld camera. The algorithm contains four parts including feature extracting and matching, projective reconstruction, self-calibration and Euclidean reconstruction, dense reconstruction. Firstly, the matched corresponding points can be obtained from the feature extracting and tracking through an image sequence, after which the projective structure will be initialed according to the geometric relations among the images. Then self-calibration will be used to determine the internal parameters of the camera in order to recover the camera motion and upgrade the projective structure to metric structure. Euclidean reconstruction of the scene will be obtained combined with an absolute yardstick, and the dense reconstruction of the scene can be obtained further with stereo matching algorithm. In the end of the thesis, the algorithm is applied to outdoor obstacle detection and three-dimensional modeling of a scene, and the results of several experiments are provided.