An Autonomous Stereo-vision-based System

A system capable of processing visual information in real-time is a common need for many arti cial vision tasks. Biological vision systems are highly interconnected systems, but it is reasonable to assume that low-level vision operations could be described in terms of independent ltering operations, such as direction detection, disparity and velocity estimation, etc. These functional operators extract speci c raw information, or feature maps, that constitute the basis for higher level processing. The homogeneous structure of the visual cortex suggests that in the human visual system the same computational mechanism is used to realize di erent functionalities. Likewise, in arti cial vision systems, a exible procedure can be used as a common processing tool. Gabor lters are a good candidate for this procedure. The local frequency spectrum of the signal computed by means of Gabor lters can be used to estimate local direction, disparity and optical ow. The choice is theoretically dictated by the fact that the Gabor representation is the best one to simultaneously represent position and frequency. The aim of this project is to design a simpli ed vision system to implement multimodal realtime vision processing using a common set of primitive operators. The Gabor representation of the visual signal is used for depth estimation, optical ow estimation, and local direction extraction. The corresponding modules produce depth, velocity and local-direction maps associated to the input images. In the design of the visual system, we are more concerned in the real-time implementation and integration of the di erent modules than in the realization of a single complex module aimed at exhaustively and completely analyze the single images. In fact, we believe that robustness and reliability of biological { and arti cial { vision systems are mostly based on the integration of the information coming from di erent types of sources rather than on an extremely specialized and precise processing.