A novel stereovision architecture for real-time obstacle detection

This paper presents a novel stereo architecture to implement a stereovision algorithm suited to obstacle detection and collision avoidance applications in real-time environments. The algorithm is derived from the spatio-frequency analysis proposed by Marr and Poggio (1979), with the matching primitives resulting from a local extremum extraction in the band-pass filtered stereopair images. A VLSI implementation is motivated both by the high processing speed required in such a real-time content and by the simplicity and regularity of the algorithm. We have therefore designed a retina-like architecture, with an in-line parallel processing unit interfaced with an on-chip photodiode matrix sensor. This stereo retina system has been succesfully simulated with real scenes images taken from a running car on a highway. A 128-pixel line CMOS retina has been designed and fabricated. First-hand experiments are positive