3-D SURFACE DESCRIPTION FROM BINOCULAR STEREO

We present a stereo vision system that attempts to achieve robustness with respect to scene characteristics, from textured outdoor scenes to environments composed of highly regular man-made objects. Unlike most stereo approaches, it integrates "area-based" and "feature-based" primitives. This allows it to take advantage of the unique attributes of each of these techniques. The area-based processing provides a dense disparity map, and the feature-based processing provides an accurate location of discontinuities. We are able to generate a disparity map that is sufficiently accurate to allow us to detect depth and surface orientation discontinuities, provided that the resolution is fine enough. We use an area-based cross correlation along with an ordering constraint and a weak surface smoothness assumption to produce an initial disparity map. Unlike other approaches, however, a match is accepted only if both views agree on a correlation peak and if this peak is strong enough. This disparity map is only a blurred version of the true one because of the smoothing introduced by the cross correlation. The problem is most acute at depth discontinuities but can be reduced by introducing the edge information: The disparity map is smoothed (subject to the constraint that the disparity at edgels is fixed) and the unsupported points removed. It is important to note that this method gives an active role to edgels parallel to the epipolar lines, whereas they are discarded in most feature-based systems. We have obtained very good results on complex scenes in different domains and have been able to locate visible surfaces, penumbral and off-the-edge areas, and depth and orientation discontinuities in the images.