Control of a camera for active vision: foveated vision, smooth tracking and saccade

In an active vision system, the information available for feedback has to be computed from images acquired on real-time. This image processing task can be seen as a degree of freedom that the designer has at hand to maximize the tracking performance, for instance one can define a region of the image where heavy processing is performed - the fovea. Based on control considerations, this paper shows how to compute the optimal size of this fovea, and presents a two-tracking mechanism which is necessary if foveated vision is to be implemented on a real environment.