Object Tracking with Stereo Vision

Task directed, active vision techniques are beginning to produce vision systems capable of providing real-time depth perception for robots. The concepts of shallow disparity filters and coarse disparity segmentation show particular promise for several reasons: they require minimal correlation search, thus they are fast; they employ redundant measurements, so they are noise tolerant; and they are not dependent on precise measurements, so they are calibration insensitive. We have extended these techniques by coarsely segmenting regions about the fixation point according to the task at hand. Relative occupancy of these regions provide cues to the approximate location of relevant features. They also provide reliable information for maintaining gaze stabilization. We have implemented these techniques on a real-time active stereo vision system to produce robust visual skills for tracking highly dynamic objects and occluding contours. These techniques are general in that they can be applied to a wide variety of objects.