Implementing and Investigating Refractoriness in LGMD Neural Networks

Collision can be threatening for animals including human beings. Thus, reliable and accurate collision perception is vital in plenty of aspects. Taking inspiration from nature, the computational methods of lobula giant movement detectors (LGMDs) identified in flying locust’s visual pathways have positively demonstrated impacts on addressing this problem. However, collision perception methods based on visual cues are still challenged by several factors in physical world including ultra-fast approaching linear velocity and noisy signals. The current visual-cue-based LGMD neural networks could show ineffectiveness or generate false positive, especially when objects approach at fast velocity and when the video signals are polluted by noises. Hence, how ultra-fast approaching object in a colliding way can be detected remains to be further improved. Neural refractoriness, also known as refractory period (RP), a common mechanism inside animals’ neural system studied for decades, though it has been considered to play a significant role in stabilising a neuron, has not been researched in the aforementioned LGMD neural networks for accurate and reliable collision perception. In this thesis, a novel method phenomenologically simulating neural refractoriness inside animals’ neural systems is proposed and is further investigated on its functionality and efficacy when it is combined with the classic LGMD1 and LGMD2 neuronal networks for collision perception. Our systematically experimental results demonstrate that, mimicking refractoriness not only enhances the LGMD1 models in terms of reliability and stability when facing ultra-fast approaching objects, but also improves its performance against visual stimuli polluted by Gaussian or Salt & Pepper noise. Potential proof of LGMD2 neural network’s reliability and its capability to adapt to cluttered physical world is also provided. This research shows that, modelling of refractoriness can be effective and benefiting in collision perception neuronal networks, and be promising to address the aforementioned challenges for collision perception