Dynamic Approaching Detection Based On Biological Visual Perception

Vision is the main sensing method for visual animals to perceive environmental information.The richness and complexity of the visual information,make the vision play an irreplaceable role in motion perception.Nowadays,building a dynamic visual system for approaching motion perception in a both reliable and efficient manner is still an open challenge.However,nature has already provided us with variously excellent paradigms to address this issue.After millions of years of evolutionary development,animals possess robust visual systems capable of motion perception to deal with challenges of living.Arthropods,in particular,have a relatively smaller number of visual neurons compared to mammals like humans,but can still smartly avoid collision against visually cluttered and dynamic environments.Understanding the arthropod visual processing neural pathways and methods thus is not only attractive to neural system researchers,but also critical in providing effective solutions for future intelligent machines.Originated from biological researches on crab’s visual systems,this thesis investigates and firstly proposes bio-plausible computational modeling of motion perception neural systems to mimic the visual processing in crab’s visual system.Specifically,the proposed models comprise approaching perception neural networks of two looming-sensitive neurons(i.e.,MLG1 and BLG1),and motion direction perception neural network of LCDCs.The contributions of this thesis are as follows:To simulate the crab’s MLG1 looming-sensitive neuronal ensemble,we take inspiration from the neuromorphic structure of the well-studied LGMD computational neural networks and proposed the first MLG1 s computational model to spatially localize the impending collision,i.e.,the looming spatial localization neural network.This model utilizes the Winner-Take-All mechanism to judge the location of the firstly approaching stimulus.A bio-plausible spike frequency adaptation mechanism has been introduced in this model to shape the selectivity of the neurons from the approaching and receding motions.To extract movement cues of wide-field motions,we propose a post-synaptic neural network of motion-sensitive neuronal ensemble,i.e.accelerating motion perception model.Within this model,we design a spike-timing based time difference encoder,which encodes the ”time-of-travel” between two successive pulses generated from two neighboring neurons.In other words,the neuronal direction selectivity is achieved by temporally and spatially comparing the information.For accelerating perception,we set an adaptive spiking-threshold based on the dynamic EPSC.Therefore,only accelerating translational motion could activate the accelerating detector.The model responses to translational accelerating motion also ft many of the explored physiological features of directional selective neurons in the lobula complex of crab(i.e.LCDCs).Together,this provides a hypothesis of motion direction and accelerating perception and produces reasonable predictions for future neurophysiological experiments.To better enhance the capability of visual perception for descending approaching motions,we further develop a bio-plausible BLG1 neural network.The pre-synaptic part of this model is identical to the MLG1s-inspired model,except for the local receptive field normalization mechanism and a feed-forward contrast inhibition pathway.The systematic experiments illustrate that the contrast normalization pathway effectively reduces the variability of the neuronal outputs.That is to say,the neural network model accommodates invariance when challenged by different contrasts.What’s more,we also design a radially extending spatial enhancement distribution to selectively enhance the signal generated from the descending approaching motions,which matches the biological feature of BLG1 neuron.The effectiveness and flexibility of the proposed motion perception neural networks have been validated by systematic and comparative experiments ranging from off-line synthetic to real-world tests,which also indicate great potential of building neuromorphic sensor strategies in autonomous insect-size robots for motion perception in a fast,reliable and low-energy manner.Note that,the explication and justification of all models in this thesis are placed within a framework for describing and comparing features in the biological behavioral and neural sciences.More broadly,we demonstrate the conceptual value of marrying artificial neural systems with biological characterization,moving beyond “black box” reproduction of an architecture to advance understanding of complex systems,such as the arthropod visual system.Furthermore,this work is expected to provide a valuable inspiration for further interdisciplinary studies.