Individuating unknown objects by combining motion and stereo
Oussama Khatib, Ramin Zabih · 1994
Unstructured environments pose a major challenge for computer vision. In such environments, objects cannot be treated as rigid bodies whose shapes are known in advance. However in many situations objects of unknown shape can be individuated by means of motion or stereo discontinuities. A motion discontinuity occurs where an element of a scene is undergoing a different motion from its neighbors; a stereo discontinuity, where an element is at a different depth from its neighbors. Such discontinuities commonly occur at the edges of objects, and thus form a vital clue for individuation. This thesis addresses the use of motion and stereo for unstructured environments, with special emphasis on the role of discontinuities. I describe a new class of algorithms for the computation of motion and stereo. The new algorithms use non-parametric local transforms as the basis for matching. Non-parametric local transforms rely on the relative ordering of local intensity values, and not on the intensity values themselves. These algorithms appear to exhibit superior accuracy compared with conventional algorithms. This is especially true near discontinuities, where most algorithms for correspondence behave poorly. I present theoretical and empirical analyses of these new algorithms, and give fast algorithms for their implementation. I also describe a real-time system that tracks moving objects of unknown shape. It relies extensively on motion and on motion discontinuities. This system functions robustly under a restricted class of camera motions (horizontal or vertical rotations around the nodal point). Finally, I present an extension to this system, which can handle arbitrary camera motions by relying on the combination of motion and stereo.