Analog-VLSI, Array-Processor Based, Bayesian, Multi-Scale Optical Flow Estimation L.T¨or¨ok, ´ A.Zar´andy Analogic and Neural Computing Systems Laboratory, Computer and Automation Research Institute Hungarian Academy of Sciences, H-1111, Budapest, Kende u. 13-17, Hungary, e-mail: torok,zarandy@sztaki.hu November 10, 2005 Abstract Optical flow (OF) estimation aims at derive a motion-vector field that characterizes motions on a video sequence of images. In this paper we propose a new multi-scale (or scale-space) algorithm that generates OF on Cellular Neural/Non-linear Network Universal Machine (CNN-UM), a general purpose analog-VLSI hardware, at resolution of 128x128 with fair accuracy and working over a speed of 100 frames per second. The performance of the hardware implementation of the proposed algorithm is measured on a standard image sequence. As far as we are concerned, this is the first time when an OF estimator hardware is tested on a practical-size standard image sequence. Keywords: Multi-scale, Optical Flow Algorithm, Analog VLSI Implementation, Cellular Neural/Non-linear Network 1 Introduction In general, a visual sensory system (VSS) can be interpreted as a system that records light in a 2 dimensional (image) plane reflected by illuminated, 3 dimensional, real-world objects. The objects of the environment that are subject to motion relative to the observer induce a motion-vector field in VSSs. The objective of Optical Flow (OF) calculations is to estimate this vector field, which describes the transformation of one image to the consecutive one. Far from being exhaustive, a review on existing OF methodologies will be given along with relation to the proposed method in sec. 2 In a simplistic approach one can classify techniques used in OF deduction depending on the basis as optical flow constraint (OFC), gray value matching (e.g. block matching (BM)), correlation (ref. [Ana89]), phase (cross correlation), higher order statistics (ref. [BJ96, ?]), or energy [Hee88, SRC98]. 1