Future Directions for Motion Detection Based on the Parallel Computational Intelligence of Insects. zyxwv Derek Abbott, Abdesselam Bouzerdoum Centre for High Performance Integrated zyxwvuts - Kamran Eshraghian Dept. of Elec. Computer and Comms Engineering Technologies & Systems (CHiPTec) Dept. of Electrical & Electronic Engineering University of Adelaide SA 5005, Australia zyxwvut dabbott@eleceng.adelaide.edu.au Abstract zyxwvutsrq The zyxwvutsrqponml visual system of insects consists of distributed neural processing, inherent parallelism andfuLy collision avoid- ance algorithms. This forms the basis for artificial vi- sion systems that exploit these computational intelligence schemes for anti-collision tasks. Insects zyxwvutsrq tend to detect mo- tion rather than images and this together with the paral- lelism in their visual architecture, leads to an efficient and compact means of collision avoidance. A family of VLSI smart microsensors that mimic the early visual processing stage in insects has been developed. The qstem employs the zyxwvutsrq ‘sman sensor’ paradigm in that the detectors and process- ing circuit7 are integrated zyxwvutsrq on one chip. The IC is ideal f o r motion detection, pUrtiCUh7rlycollision avoidance rash, as it essentially detects the speed, bearing and time-to-impact of a moving object. Fuzzy algorithms may then be employed for decision making. The Horridge model for insect vision has been directly mapped into VU1 and therefore the IC truly exploits the beauty of nature in that the insect eye is so compact with parallel processing, enabling compact mo- tion detection without the computational overhead of intert- sive imaging, full image extraction and interpretation. This world-first has exciting applications in areas such as anti- collision for automobiles and autonomous robots. The sta- tus and future directions of this work are outlined 1. Introduction For collision avoidance tasks, where full imaging is un- eccessary, simple motion detection of object boundaries offers an efficient solution. A world-first single-chip de- vice, based on insect vision principles, bas been developed [ 1, 2, 3, 4, 5, 61 that outputs the time-to-impact, bearing Edith Cowan University Joondalup WA 6027 Australia k.eshraghian @cowan.edu.au and velocity of a detected object. The processing power of a commercial microcontroller is then sufficient for making decisions based only on such simple variables. The insect vision model we have adopted is that of Hor- ridge [7, zyxwvut 81, which is named the ’template model.’ This model is motivated by the desire to produce signals which can be readily interpreted by digital systems, and hence readily lends itself for mapping onto a VLSI chip. The edge of an object presenting a difference in contrast with the background, and moving in front of an array of recep- tors, elicits distinctive pattems of contrast changes which are consistent with the direction of motion. The receptor outputs are sampled and compared with their previous val- ues, yielding signals which locally indicate an increase, de- crease. or no change in contrast. The combination of two adjacent receptor responses at consecutive sampling times form a ‘template,’ and hence, since there are three pos- sible receptor responses. there are 81 possible templates. The VLSI implementation and detailed description of this scheme is described elsewhere. In this paper, we firstly review the template model and compare this with the biological insect eye architecture. We. then proceed to discuss some of the implementation prob- lems, developments that have occurred to solve them and the future vision. 2. Overview of insect vision Insects, compared to humans, possess a relatively sim- ple visual system, yet are capable of performing complex visual tasks The insect visual system has a highly parallel smcture - the visual ganglia in the optic lobes are organ- ised into columns and strata The lamina is the first optic ganglion and contams a large number of identical channels The photoreceptors. in the rema, sample the visual field and 244 0-8186-8215-9/97 $10.00 zyx 0 1997 IEEE