John Woodfill, Ramin Zabih
We describe an algorithm for tracking an unknown object in natural scenes. We require that the object’s approximate initial location be available, and further assume that the it can be distinguished from the background by motion or stereo. No constraint is placed on the object’s shape, other than that it not change too rapidly from frame to frame. Objects are tracked and segmented cross-temporally using a massively parallel bottom-up approach. Our algorithm has been implemented on a Connection Machine, and runs in real time (15-20 frames per second).