Ian Douglas Horswill, Rodney Allen Brooks
We describe a system that approaches and follows arbitrary moving objects in real time using vision as its only sense. The system uses multiple simple vision computations which, although individually unreliable, complement each other in a manner mediated by a situated control network. The objects can move over a wide variety of backgrounds including those with strong secondary reflections from light sources. Previously unseen objects can be tracked against backgrounds that include other moving objects. Computations are carried out in image coordinates at roughly 5 frames per second on a Lisp machine. The camera need not be calibrated or aligned well, and the system can tolerate a wide range of dynamically changing actuator response characteristics.