A new dense omnidirectional stereo matching approach

This paper presents a new method for matching omnidirectional stereo images. The main idea of the proposed method consists in exploiting the temporal information for matching the stereo images acquired at the current time by involving the stereo images acquired at the preceding time. Three pairs are considered in the matching process, i.e., the current, the left temporally adjacent and the right temporally adjacent omnidirectional stereo images. All the three stereo pairs are used together in order to overcome the problem of blind spots near the epipoles. The proposed approach is achieved in three main steps. The first step performs the spherical rectification on the three input pairs. In the second step, we apply the proposed dynamic programming-based stereo matching algorithm on the resulting three rectified pairs. Disparity ranges for each scanline are computed from the v-disparity maps generated based on the disparity maps computed at the preceding time. This reduces the matching candidates and speeds up the matching process. In the final step, we combine the three disparity maps into a single 360° disparity map. The proposed method has been tested on real omnidirectional images and the results provided are promising.