The low-level representation of visual motion utilized by biological organisms from insects to primates is fundamentally different from that used in conventional computer vision systems. Rather than an optical flow vector field, banks of nonlinear spatio-temporal frequency tuned filters are used, giving rise to a representation which naturally supports transparent motion and shear at occlusion boundaries, and which facilitates the solution of the aperture problem. Because of the amount of parallel computation required to detect motion in this fashion, a biological representation of motion requires efficient implementation of elementary motion detectors. In this paper we describe the low-power continuous-time analog VLSI implementation of a biologically-inspired visual motion sensor. This sensor, based on the motion energy algorithm of Adelson and Bergen, is a simple hardware model of the motion response of a primate cortical complex cell but can also be shown to be equivalent to the Reichardt model of insect motion detection. Characterization results show that this spatio-temporal frequency tuned sensor can discriminate the direction of motion of a sinusoidal grating down to less than 5% contrast, and over more than an order of magnitude in velocity. In addition to providing a real-time hardware biological model for investigation of spatial motion integration algorithms, this sensor will be a fundamental building block for experimentation into biologically-inspired visual navigation architectures for underwater, airborne, and land-based autonomous robots.
Charles M. Higgins and Sudhir Korrapati, "An Analog VLSI Motion Energy Sensor Based on the Adelson-Bergen Algorithm," in Proceedings of the International ICSC Symposium on Biologically-Inspired Systems, Wollongong, Australia, December 12-15, 2000.