We introduce a highly parallel continuous-time biologically inspired computational architecture for detection and spatial integration of visual motion. This architecture begins with two-dimensional image transduction and signal conditioning, performs low-level motion detection with a number of parallel elementary motion detector arrays, and then spatially integrates the low-level motion detectors to synthesize units sensitive to complex wide-field patterns of visual motion. This high-level architecture is based on the general organizing principles of visual motion processing common to organisms from insects to primates. A successful hardware implementation in low-power custom neuromorphic VLSI is described, demonstrating the ability of the system to selectively respond to certain motion patterns, such as those that might be encountered in self-motion, at the exclusion of others. This hardware system is designed to become a compact visual system for small autonomous robots with applications in visual navigation, self-motion estimation, and target tracking. Extensions of this architecture to include simultaneous binocular disparity computation are discussed. Finally, a revised hardware implementation based specifically on models of fly visual motion processing is outlined which will improve motion detector low-contrast response and allow clear representation of transparent motion and occlusion boundaries, making the system more powerful for real-world applications.
Charles M. Higgins and Shaikh A. Shams, "A Neuromorphic Vision Processor for Spatial Integration of Optical Flow", (abstract) in Proceedings of the Fifth International Conference on Cognitive and Neural Systems, Boston, Massachussetts, May 30-June 2, 2001.