On computer stereo vision with wire frame models.

Abstract : This report treats two major subproblems in computer stereo vision -- that of reconstructing 3-D scenes from stereo sets of images and automatic recognition of 3-D scenes. Two techniques are presented for scene reconstruction. The first, or multiple view method, utilizes the combined information from bulk correlation and three or more stereo images to construct three-dimensional edge features or structures. The structure is obtained by projecting into space a piecewise-linear representation of intensity edges obtained from one of the images. The second technique utilizes a narrow angle pair (2-3 degrees) of images and symbolic correlation to ensure matching reliability and efficiency in the construction of edge depth maps of scenes. A new technique of dynamic smoothing of edge contours is presented which permits accurate triangulation at narrow viewing angles, while preserving the integrity of sharp corners. Also, two new techniques are presented for piecewise approximation of 3-D and 2-D digital contours with circular arcs.