基于仿生运动能量特征的光流计算芯片设计综述

A review on chip systems for optical flow computing based on biological motion energy features

简述了基于仿生视觉运动能量特征进行光流估计的原理,分析了已有的各种模型和算法,总结了以往基于运动能量特征的光流计算芯片设计的相关工作,其中重点介绍了一款基于全数字电路处理的仿生光流估计片上系统。该仿生系统从算法理论和硬件电路两方面进行协同优化设计。算法流程包括纹理增强,基于空间-时间联合滤波的运动能量特征提取,以及速度合成3个阶段,分别对应模拟了人类视觉系统中视网膜、视皮层V1区、视皮层V5/MT区与运动感知相关的各种神经元功能。为实时实现该算法,对应的片上系统硬件架构采用了帧级、像素流级和电路级3级流水线机制,并包含多个并行处理阵列以加快处理速度。阵列处理单元电路进行了优化设计,以减少硬件成本开销。基于Zynq-7020低成本FPGA平台实现了该仿生光流计算片上系统原型,每秒可实时估算30帧、每帧160×120个像素点的运动速度,平均误差小于0.5个像素。最后简单分析了限制目前基于仿生运动能量特征的光流计算系统准确度的关键因素,介绍了初步的解决方案,给出了未来该领域研究发展的方向。

This review paper briefly explains the principle of optical flow estimation based on biological visual motion energy features,analyzes published relevant models and algorithms,and summarizes research works on optical flow computing chips based on such biological features.Among those works,a system-on-chip(SoC)design with fully digital circuits processing is especially focused on,which is algorithm/hardware co-optimized.Its algorithm flow consists of three stages:texture enhancement,motion energy feature extraction by spatiotemporal filtering,and motion velocity synthesizing,corresponding to retina,cortex V1 and cortex V5/MT regions relevant to motion perception in human visual system,respectively.To execute the algorithm in real time,the SoC hardware adopts multiple levels of pipeline(i.e.,frame-level,pixel-stream-level and circuitslevel),as well as multiple parallel processing arrays to accelerate the processing speed.The circuits of array processing units are optimized to reduce hardware resource costs.An FPGA prototype of this biological SoC implemented on a low-cost Zynq-7020 platform is demonstrated,which can estimate motion velocities on 160×120 pixel locations at a speed of 30 frames per second(fps)in real time,with a small average End-point Error(EE)below 0.5 pixel.Finally,this paper reveals bottlenecks that limit the accuracy of optical flows computed by the currently available biological systems,as well as introduces preliminary solutions to guide the future directions of researches in this field.