Modelling of bio-inspired vision system for velocity estimation.

Abstract

Although motion processing in insects has been extensively studied for over almost 40 years, velocity detection in insects and how the insect brain computes the velocity of a moving feature, independent of its size or contrast, is a major enigma that remains unsolved. This study examines the accuracy of velocity estimation using two biologi-cally inspired models of motion detection, (i) the Horridge template model and (ii) the Reichardt correlator model. Various extensions and enhancements of these models are implemented with the goal of achieving robust velocity measurements. The template model is one of the prominant models of motion detection, which was proposed by Horridge in 1990. This Thesis further extends this model with the aim of improving accuracy in velocity detection using chrominance as well as luminance channels with various error checking mechanisms using different stimuli. Then the template model response is compared with Dror's elaborated Reichardt model and electro-physiological experimental results obtained from the fly visual system using similar stimuli in each case. A modified Reichardt model is shown to give a more similar response to that of fly neurons. In order to improve velocity performance of the Reichardt model, it is necessary to reduce contrast dependance of the correlator response as well as to make it independent of the structure of the visual stimuli. With this aim, the Reichardt model is then further elaborated to include contrast adaptation by a feedback adaptive mechanism and a clear reduction in contrast dependency is demonstrated. The deviation of the correlator response depending on the stimulus is termed as pattern noise. To reduce this pattern noise, the Reichardt correlator model is further extended to implement compressive non-linearity or saturation. It is seen that saturation has a profound effect on the shape of the pattern noise. Further studies on pattern noise is then performed in this Thesis using different stimuli at different speeds and contrasts. Work carried out in this Thesis on the affect of various receptive field shapes on the pattern noise reveals circular sampled arrays reduce pattern noise and hence, based on this result, a small 16 pixel yaw sensor using our elaborated model is built that shows promising performance with various potential applications.