Geometrically Constrained RANSAC for Stereo Image Registration in Presence of High Ambiguity in Feature Correspondence

An approach for registration of sparse feature sets detected in two stereo image pairs taken from two different views is proposed. Analogously to many existing image registration approaches, our method consists of initial matching of features using local descriptors followed by a RANSAC-based procedure. The proposed approach is especially suitable for cases where there is a high percentage of false initial matches. The strategy proposed in this paper is to modify the hypothesis generation step of the basic RANSAC approach by performing a multiplestep procedure which uses geometric constraints in order to reduce the probability of false correspondences in the hypothesis. The algorithm needs approximate information about the relative camera pose between the two views obtained e.g. by odometry. However, the uncertainty of this information is allowed to be rather high. The presented technique is evaluated using both synthetic data and real data obtained by a stereo camera system.