Motion segmentation and local structure

The author presents a means of segmenting independently moving objects from rigid backgrounds and recovering the orientations of the normals of local planar surface patches. The camera motion can be arbitrary with the proviso that the independent image motion violates the epipolar constraint. The method relies on describing the motion of points in the world in terms of their angular velocity relative to the camera. The method is crucially dependent on estimating camera rotation around image directions. An assumption about the likely motion of automatic guided vehicles is used to resolve the two-fold ambiguity in the estimate of the normal of a local planar patch. This allows local scene structure and camera rotation to be estimated. Motion segmentation is demonstrated on a real image sequence. >