A preliminary investigation of the problem of determining ego- and object motions from stereo

The following problem is addressed: given a stereo rig moving in an unknown environment where an arbitrary number of rigid mobile objects might be present, it is required to segment the various mobile objects from the background and determine the motions of the stereo rig and of the mobile objects. The technique proposed is based on matching 3-D tokens computed from stereo at different time instants. The matching uses the standard paradigm of hypothesis prediction and testing and exploits the constraint that objects and environment are rigid. The uncertainty about the position and velocity of the tokens is explicitly taken into account in the matching process through an extended Kalman filter. Preliminary results are presented