Research on Brain-inspired Vision Based on Dynamic Vision Sensor Cameras

Abstract

In this paper, we explored many applications of brain-inspired vision, in which the data was acquired based on a dynamic vision sensor camera, that is PSEE300EVK from French. Specifically, we explored the following three aspects: (1) Converting large-scale artificial convolution neural network into spiking neural network, which can process large-scale datasets and save network resources without dropping much precision. We proposed reliable solutions for the difference between the two networks, and it can be generalized to other deep network transformations. (2) Recognizing pedestrians and vehicles spiking data flow in the autopilot scenario. Specifically, we transformed Cityscapes dataset into two modes spiking data, one called event processing mode, another called contrast detection mode. (3) Constructing a structured light 3D acquisition system and 3D image recognition algorithm based on the PSE300EVK camera. Tests show that the algorithm used in this paper can effectively reduce the error between the deep artificial convolutional neural network and the deep spiking convolutional neural network, and it has good ability of generalization. In this article we push forward application of brain-inspired vision in actual scenes and look forward the further development, besides, it helps deploy power-efficient neuromorphic spiking neuron chips in embedded applications.