A Computational Investigation of Target Detection and Tracking by Insect Small Target Motion Detector Neurons

Abstract

Vision is an important sensory modality for many animals and a useful source of information for artificial systems. For animals, effective behavioural responses require relevant features to be extracted from complex visual environments quickly and robustly. The visual processing algorithms utilised by animals have been optimised through evolutionary processes for not only performance, but also the size of the required “hardware”, and energy efficiency. Understanding these evolved algorithms may reveal useful algorithms for artificial visual systems. The visual systems of many animal species have been investigated. Among these, insects offer the advantage of a relatively simple and experimentally accessible brain. Neurons which respond selectively to small moving targets have been identified in several insect species. These neurons, named small target motion detectors (STMDs), are sensitive to target contrast and have tuned responses to target size and velocity. STMDs in dragonflies have complex, nonlinear responses including selective attention and predictive enhancement of responses to targets moving on continuous trajectories. The properties of STMDs give rise to two complementary questions, on which I have focussed: how is the spatiotemporal signature of a small moving target differentiated from distracting features in a moving naturalistic background; and, how are the responses of individual STMDs combined to produce an adequate representation of the target’s position or trajectory. I investigated these using computational modelling based partly on an existing model of insect small target motion detection, the elementary small target motion detector (ESTMD) model. I evaluated the suitability of the ESTMD model for guiding aerial pursuits using footage of a drone in flight taken by another drone in flight. The suburban environment forming the backdrop for these flights contained many features which, during smooth turns, produced false positives. These experiments showed that using a threshold on the ESTMD model outputs to differentiate target-related responses from those generated by the background was not sufficient for guidance of aerial pursuits involving smooth turns. To explore whether spatiotemporal correlations in the position of the target could be exploited to differentiate between false positives and true positives, I implemented a particle filter with a salience measure representing selective attention. I demonstrated that, in simulated pursuit scenarios, this state estimation filter improved performance in pursuing small moving targets relative to an existing bio-inspired tracking mechanism. I investigated how fast, non-linear adaptation occurring early in the insect visual system, which forms part of the ESTMD model, affects target detection against cluttered backgrounds. I carried out extensive comparisons between a set of alternative filtering mechanisms via simulation. I found that the non-linear adaptation mechanism offered a comparative advantage over the tested alternatives when the background was moving repetitively, but not otherwise. Lastly, I investigated how effectively target trajectories can be encoded in the responses of a population of STMDs. Using information theoretic measures and a novel modelling framework, I quantified how the size and arrangements affect the encoding of ethologically relevant target trajectories. These results show that the size and density of the STMD mosaic reflects trade-offs between efficiency and encoding performance.