Design Of Photon Deconvo-Lution System For Optical Flow Estimation Of Event-Based Camera

Since traditional frame-based cameras fail in high-speed motion or low-light environments,the optical flow estimation of event-based cameras with high temporal resolution and high sensitivity is widely studied to make up for the shortcomings of frame-based cameras.To cope with the complex and changeable road environment,the optical flow estimation architecture of the event-based camera often takes a long time to output the forecast results,which is not conducive to the on-board computer using the optical flow estimation results to make a quick response to dangerous situations in autonomous driving.The traditional optical flow estimation architecture is based on electronic computing,but electronic computing is difficult to further improve the computing speed due to the limitations of transistor size,von Neumann architecture,electronic components defects and other limitations,and the introduction of high frequency,high bandwidth,high parallel ability and low power consumption advantages of optical computing,can greatly improve the speed of optical flow estimation.Reduce power consumption to achieve speed and energy efficiency of the optical flow estimation architecture.The main research contents of this paper are as follows:1.A photon deconvolution acceleration system and its data preprocessing algorithm are designed.The system consists of a light source with a frequency of 193.1THz and a wavelength of 1550nm,an optical beam splitter,a light source modulator,MZI,a phase shift,an optical balance detector and other photonic components.For the data processed by the preprocessing algorithm,the deconvolution results of 16 different deconvolution cores can be calculated at one time.Moreover,the system adopts the idea that the phase characteristics of light represent the positive and negative data,and can correctly represent the positive and negative values of the output results.The system has the characteristics of fast speed and low energy consumption.Compared with the deconvolution operation based on electronic calculation,the calculation speed is about 1000 times faster in theory.2.An optical flow estimation neural network architecture of eventbased camera embedded with photon deconvolution acceleration system is designed.The architecture consists of four layers of pulse neural network,two layers of residual,three layers of deconvolution and four photon deconvolution acceleration systems.The photon deconvolution system is used to replace the deconvolution operation which occupies 11%of the time of the Spiking-FlowNet architecture,so the operation speed of the optical flow estimation is accelerated.Compared with the original architecture,the architecture is smaller in size,more energy efficient,and the calculation speed can be improved by 11%.The AEE of optical flow estimation index is 1.03,slightly better than that of the original SpikingFlowNet architecture,which is 1.07.In addition,the influence of errors of optical beam splitter and phase shifter was considered.Compared with the 13.3%increase of architecture AEE after direct mapping,architecture AEE only increased by 4.7%(1.078)after joining the training process,which is like the original architecture.