Continuous Spiking Recognition Method Based On Dynamic Vision Sensor

Brain vision system shows stronger abstract and association ability than computer vision system.As the third generation of artificial neural network,the brain-inspired spiking neural network shows better biomimetic and stronger ability of information expression by simulating human brain and simulating how neurons connect and transfer information.Spiking neurons are driven by asynchronous and discrete spikes.Dynamic vision sensor takes such event-driven way to simulate sensory neurons from vision system and thus realizes a more bionic,stable and efficient way for capturing spike signals.The cognition process of brain is continuous.This is closer to online continuous spiking recognition which is based on each single spike,and there is a big difference between this and offline discontinuous spiking recognition which is based on the whole spike train.To further simulate brain vision system and take advantage of spiking neural networks,in this paper,we propose a novel model for offline discontinuous and online continuous spiking recognition.This model is based on vision hierarchical model and asynchronous spikes with low latency from dynamic vision sensor,and achieves better performance and stability.The contribution includes:1.Propose a new feature extraction method towards discontinuous spiking recognition,based on vision hierarchical model.This method takes Gabor filter as neuron connection way and extract feature spike trains by neuron membrane.And a coding method based on the neuron membrane potential attenuation is proposed,which obeys the feature distribution and achieves better performance.2.Propose a new event-driven feature extraction method towards continuous spiking recognition,by combing neural firing and lateral inhibition with vision hierarchical model.This model achieves both better performance and faster recognition.3.Propose a spiking learning rule on neuron spike frequency mapping function and Tempotron learning rule.This method takes spike frequency mapping function to show better anti-jamming ability and takes Tempotron method as regular term to make it more stable.This method achieves better performance and stability,especially on small dataset,iteration stability and anti-noise performance.Meanwhile its variants based on segmented weight enhancement achieves faster recognition speed towards continuous spiking recognition.4.Design and develop an online gesture recognition system,including online recognition and offline analysis features.This system is designed to verify the continuous recognition effect of the proposed online or offline spiking recognition model in real tasks.