An analog VLSI motion sensor based on the fly visual system

Author: Harrison, Reid R.

Year: 2000

Degree: Dissertation (Ph.D.)

Advisor: Koch, Christof

Committee Members: Dickinson, Michael H.; Goodman, Rodney M.; Laurent, Gilles J.; Perona, Pietro

Option: Computation and Neural Systems

DOI: 10.7907/TSDS-SJ16

Abstract

Vision is a vitally important sense for flying insects. Over half of the 350,000 neurons in the housefly's brain are believed to have some role in visual processing. Flies use visual motion cues to navigate through turbulent air, avoid obstacles, chase other flies, and land safely. Much is known about the neural circuitry that extracts motion information from retinal light intensity signals. This dissertation describes the development and testing of integrated silicon sensors that estimate visual motion using architectures derived from the neurophysiology of the fly optic lobe. We built VLSI systems incorporating light sensors and information processing circuits side by side on the same chip. These continuous-time analog CMOS circuits operate in the weak inversion (subthreshold) regime to match biological time constants and achieve sub-milliwatt power dissipation. Detailed characterization showed our sensor to be an accurate implementation of the Hassenstein-Reichardt motion detector model, originally developed to describe insect visual responses. We developed a novel test paradigm using stimuli with natural image statistics and spatiotemporal noise to evaluate the sensor's robustness. Our sensors were able to discriminate motion direction using naturalistic stimuli in noisy conditions (SNR < 1). Information theoretic techniques were used to measure the ability of our sensor to encode time-varying image velocity. Coding efficiency was quantified and compared with results from motion-sensitive neurons in the fly. The silicon system was tested in the context of a visually guided behavior-the optomotor stabilization response. Direct comparisons with the fly were made in real-time, closed-loop control experiments. A circuit architecture was developed to model the biophysical properties of wide-field motion-sensitive neurons in greater detail. This gave our sensor array nonlinear spatial integration properties that decreased sensitivity to gaps in the optic flow field. Finally, we investigated the issue of sensory fusion and explored a circuit that could assist in the integration of visual motion sensors with other sensors, such as angular rate gyroscopes.

Files