Abstract
This paper describes a stereo matching algorithm for occluding contours in a cluttered background. It is generally difficult to match occluding contours in stereo images because the edge points of occluding contours may not be extracted or the contrasts of the corresponding edge points are not always similar.
The candidate region of an occluding contour is determined from the occluded region using a geometric constraint. In the first stage, edge points are extracted in the candidate region. For both the left small region (called a window) and the right window of an edge point, the possible regions in the other image are searched for the matching window respectively. If both the left and right windows are well matched, the edge point is determined to be the initial point on the occluding contour. In the second stage, the remaining part of the occluding contour is acquired. The disparities of the points in the remaining part are interpolated by using those of the known occluding contour points. A candidate point is supposed to be on the occluding contour, and then the degree of the matching is computed with the corresponding point, whose position is determined by the disparity, in the other image. The occluding contour is acquired by connecting the candidate points which optimize the criterion function in terms of the degree of matching and the smoothness of the occluding contour. Experiments have been executed for artificial and real images to verify this algorithm.
