Cutting Edge: The key to artificial vision is staring us in the face

If you want to know how a locust avoids high-speed collisions, take it to a movie. Roger Santer reports.

Animals' visual systems are incredibly good at extracting useful information from a very complex visual world. But humans' artificial "seeing machines" tend to fall short of the prowess achieved by many animals. They are unable to access the appropriate cues as quickly, easily and cheaply.

This should not surprise us. Good design takes time to perfect, and time is one thing Mother Nature has had. Her visual solutions have been shaped by millions of years of evolution and have reached a degree of perfection that human engineers can only dream of. The solution may be to copy what works in nature - the process of biomimetics.

Research at Newcastle University is concerned with one particular facet of vision, the detection of "looming". In essence, a looming stimulus is one that appears to approach the observer directly. By being able to identify such stimuli successfully, an organism may avoid collision or capture by a predator. By understanding how that organism achieves this, we may be able to produce an artificial system that can do the same. As inspiration we have taken the locust, Locusta migratoria . Far from being simple creatures, locusts operate within the same complex world as we do. Their compound eyes, along with those of other insects, are particularly special. They have an incredible temporal resolution - a fly can resolve 133 separate flashes of light in a second - that allows them to perceive even the most rapid of movements as if they occur in slow motion. A vast proportion of their brain is dedicated to dealing with visual input and coding it into the information needed to follow a mate, avoid collision or judge distance.

Research at Newcastle was pioneered, and is still led, by Claire Rind and Peter Simmons. As biologists, they were intrigued by how the locust nervous system can identify the visual "features" defining a potentially colliding object. When they began, a particular locust neuron, the DCMD (descending contralateral movement detector), was known in some detail. They were able to confirm that this nerve cell was responsible for collision detection. In addition, they were able to identify the particular visual cues that alert the locust to an imminent collision. Part of their research involved showing the locust clips from the film Star Wars while monitoring the reaction of its nervous system.

From these beginnings, research continued until we had a firm concept of the precise neural architecture of the locust's collision-detecting system. At this stage, it became clear that a synthetic version of this system could have a potential application as an artificial visual sensor. With these aims in mind, Rind began a variety of collaborations, most notably with researchers at the Institute of Neuroinformatics in Zurich, that led to the creation of a prototype computational model of the locust collision-detecting system. When used to control the movements of a small mobile robot, this model showed remarkable success in avoiding collisions with static objects.

Although this research is still in its infancy, it shows promise. With a firm concept of how the collision-detecting system functions, we can begin to look at how the locust takes this processed information and converts it into a neural command for evasive behaviour. We may even be able to apply these findings to our biomimetic work and, in doing so, gain an understanding of how the complete collision-avoidance pathway, from detection to evasion, functions.

Roger Santer is a postgraduate student, supervised by Claire Rind and Peter Simmons, in the department of neuroscience at Newcastle University.