A spiking neural network architecture for visual motion estimation

Abstract

Current interest in neuromorphic computing continues to drive development of sensors and hardware for spike-based computation. Here we describe a hierarchical architecture for visual motion estimation which uses a spiking neural network to exploit the sparse high temporal resolution data provided by neuromorphic vision sensors. Although spike-based computation differs from traditional computer vision approaches, our architecture is similar in principle to the canonical Lucas-Kanade algorithm. Output spikes from the architecture represent the direction of motion to the nearest 45 degrees, and the speed within a factor of √2 over the range 0.02 to 0.27 pixels/ms. © 2013 IEEE.