In this paper we present a hybrid stereo matching algorithm for short baseline stereo. With this coarse-to-fine algorithm, the stereo matching is first guided by the 2D affine transform and then completed suing correlation based on the intensity, and nonparametric transforms including the rank transform and the census transform. With only three known or detected reference points in the stereo images, the 2D affine transform can be recovered and used to model the stereo system providing the approximate disparity information for any other point. To deal with the perspective, a vote, which is based on the correlation of the intensity and the nonparametric transform, is carried out for precise stereo matching. Since it is guided by the 2D affine transform and limited to small neighbor windows, the stereo matching algorithm is especially efficient in situations where reference points are available and either the image scene is nearly planar or the scene is sufficiently far away form the cameras compared with the baseline. Experimental results proving the performance of our algorithm are presented.
展开▼