Voting-based grouping and interpretation of visual motion

A main difficulty for estimating camera and scene geometry from a set of point correspondences is caused by the presence of false matches and independently moving objects. Given two images, after obtaining the matching points, they are usually filtered by an outlier rejection step before being used to solve for epipolar geometry and 3D structure estimation. In the presence of moving objects, image registration becomes a more challenging problem, as the matching and registration phases become interdependent. We propose a novel approach that decouples the above operations, allowing for explicit and separate handling of matching, outlier rejection, grouping, and recovery of camera and scene structure. The method is based on a voting-based computational framework for motion analysis; it determines an accurate representation, in terms of dense velocities, segmented motion regions and boundaries, by using only the smoothness of image motion, followed by the extraction of scene and camera 3D geometry.