In this thesis a hardware motion processor is presented that is sensitive to complex spatial patterns of motion in the visual field; for example, patterns of expansion, contraction, and rotation which might be encountered in the self-motion of a robotic system. Biological motion processing strategies use multiple stages of simple parallel processors. However, the extent of pixel-parallel focal plane image processing is limited by pixel area and imager fill factor. Addition of more processing dramatically reduces the number of pixels on a reasonable-sized die. An obvious solution to this dilemma is the use of a multiple-chip system. The motion computation is split into more than one chip: a photosensitive sender, motion processing transceiver chips and an integrating receiver. An asynchronous digital inter-chip communication protocol is used in communication between analog VLSI chips designed on neuromorphic principles. This multichip motion processor retains the primary advantages of focal plane neuromorphic image processing: low power consumption, continuous-time operation, and small size. The sender chip detects moving edges in the image focused onto it and transmits that information to the transceivers. Using the position and timing information of edges detected by the sender chip, each pixel of the transceiver chip computes the local one-dimensional velocity of moving edges and transmits that information to the receiver chip. The receiver chip spatially integrates the motion information coming from the transceiver chips to produce sensitivity to patterns of optical flow. Using EPROMs on the way from sender to transceivers, the visual field is rotated such that each identical transceiver chip detects a different direction of motion. The information from the transceivers is combined by a routing processor in prearranged patterns and transmitted to the receiver chip. This multichip neuromorphic motion processor is ideal for sophisticated, real-time onboard sensors for autonomous robotics applications.
Shaikh Arif Shams, "A Multichip Neuromorphic Motion Processor for Extracting Egomotion Information," MS thesis (advisor: Charles M. Higgins), Department of Electrical and Computer Engineering, The University of Arizona, September 2000.