A multistage stereo method giving priority to reliable matching

The main problem with stereo vision is the method of searching for corresponding points in the left and right images. The correspondence problem becomes especially difficult in the case of complex scenes because many corresponding candidates are found for a weak feature point. This paper proposes a multistage stereo method which determines the more reliable matching pairs first and later those less reliable, with the latter determined based on the first result. A feature point is a zero-crossing point (ZC). If the contrast is high at a ZC, then the position and the direction of the ZC are clearly determined and the matching result is also reliable. Therefore, the proposed method determines first the correspondence of ZCs with higher contrast. For each ZC in the left image, the corresponding candidates in the right image are those with a similar contrast and direction. For ZCs in a small region in the left image, a histogram of the disparity of all the matching pair candidates is obtained and the disparity of the region is determined by the profile of the histogram. Once the disparity of the window is obtained, the disparity of each ZC in the region is determined. After the forementioned processing is finished, a similar processing is repeated for ZCs with lower contrasts. Experiments with complicated indoor scenes proved that the proposed method is better than a conventional single-stage method.