Visual motion perception plays a vital role in the process of behavioral computations that are performed during the interaction of organisms with their environment. In the early stages of motion detection, visual information is processed by a large number of elementary motion detectors to obtain a representation of the visual field in terms of local motion vectors. In this work, we present analog VLSI implementations of motion detection algorithms that are based not only on biological models but also on the computational properties of motion perception. First, we describe monolithic implementations of hysteretic winner-take-all and nonlinear-differentiator based algorithms. These compact elementary motion detector models can reliably be used to obtain high resolution sensors. Second, we explain multi-chip implementations of biomimetic intensity-based models, namely Adelson-Bergen, Hassenstein-Reichardt, and Barlow-Levick models. By employing a modular strategy, these algorithms are successfully implemented without much sacrifice of the fill factor in the front-end chip. In addition, we describe an obstacle avoidance algorithm that is realized by incorporating a multi-chip version of the Adelson-Bergen algorithm with centering behavior and time-to-collision computation. In this way, the overall system can successfully acquire clues about the structure of its environment so that collisions can be effectively avoided. This system might be employed in building a robot that can navigate in complex cluttered environments.
Erhan Ozalevli, "Analog VLSI Implementations of Visual Motion Sensors and a Neuromorphic Obstacle Avoidance System," MS thesis (advisor: Charles M. Higgins), Department of Electrical and Computer Engineering, The University of Arizona, May 2003.