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

摘要

An approach for registration of sparse feature setsrndetected in two stereo image pairs taken from two different viewsrnis proposed. Analogously to many existing image registrationrnapproaches, our method consists of initial matching of featuresrnusing local descriptors followed by a RANSAC-based procedure.rnThe proposed approach is especially suitable for cases wherernthere is a high percentage of false initial matches. The strategyrnproposed in this paper is to modify the hypothesis generationrnstep of the basic RANSAC approach by performing a multiplesteprnprocedure which uses geometric constraints in order tornreduce the probability of false correspondences in the hypothesis.rnThe algorithm needs approximate information about the relativerncamera pose between the two views obtained e.g. by odometry.rnHowever, the uncertainty of this information is allowed to bernrather high. The presented technique is evaluated using bothrnsynthetic data and real data obtained by a stereo camera system.