Determining the dominant plane from uncalibrated stereo vision by a robust and convergent iterative approach without correspondence

A robust, iterative approach is introduced for finding the dominant plane in a scene using stereo vision. Neither camera calibration nor stereo correspondence is required. Iterative two-step methods have been applied to various problems in computer vision, but only recently Cohen has formalized a framework that guarantees (at least local) convergence of such methods. In this paper we adopt this framework to the specific case in which the global step uses tentative matches to estimate the planar projectivity and the local step attempts to solve the stereo correspondence. The points in the second image are only used indirectly in the matching; instead we match to auxiliary points which lie on the line joining the transformed image points and their closest points in the second image. To assure convergence, weights for the auxiliary points are computed corresponding to a selected robust function, making our approach robust to both mismatches and points protruding from the dominant plane.