Improved temporal correspondences in stereo-vision by RANSAC

Determining the image correspondences of scene points in two or more views remain elusive especially for large baselines and (or) rotations. Assuming a calibrated stereo rig, we propose two novel RANSAC-based algorithms to establish temporal correspondences of triplet and quadruplet matches in two stereo pairs robustly. We devise an efficient technique for the absolute orientation problem with minimum three matches [B. Horn et al., 1987]. This drastically reduces the number of random samplings, for a given probability of selecting no outlier in at least one set, compared to the application of the 7-/8-point algorithms [Z. Zhang et al., 1995, P. Torr and D. Murray, 1997]. Within the general RANSAC paradigm, a low-cost selective random sampling scheme is incorporated to further enhance the robustness of the algorithm that utilizes quadruplet matches.