Representations for Active Vision

Abstract

As the field of Computational Vision matures, more efforts are devoted to vision systems that are active and need to interact with their environment in real time. A prerequisite for integrating Vision and Action is the development of a set of representations of the visual system's space-time, where space includes the system itself. Thus we are faced with the problem of studying the nature of appropriate representations and also with the computational task of acquiring them in a robust manner and in real time. Both of these problems are addressed in this paper from a computational point of view. In particular, we study representations needed by active visual systems in order to understand their self-motion and the structure of their environment. The representations are of less metric information content than the ones traditionally used, including depth, surface normals, curvature and 3-D metric values for the parameters of rigid motion, etc.; but they are rich enough to allow the system to perform a large number of actions. These representations, indexed in image coordinates, are the direction of translation and the direction of rotation for the case of motion and a monotonic function of the depth value in the case of shape description. Their advantage comes from the fact that they can be computed from minimal and well-defined input (flow or disparity values along image gradients), as opposed to the traditional ones which require image correspondence or the utilization of assumptions about the environment.