图像序列的光流估计理论在机器视觉领域已被提出多年, 但算法的高计算复杂度限制了其在工业领域的应用。为了满足应用的实时性要求, 阐述了一种光流实时估计的实现方法。为了提高算法精度及减少FPGA片内资源消耗, 对L&K光流计算方法进行改进。首先, 通过设计两层光流计算架构来提高精度。针对在此过程中出现的外部存储器读写速率不够的问题, 提出一次读取同时分层缓存、分时计算的方法。考虑到两层光流在计算过程中的迭代关联性, 设计了满足要求的外部存储器数据读出顺序表; 然后, 针对卷积运算资源消耗大的问题, 设计了新的卷积权重函数, 能够将卷积计算量降低73%, 从而节省了大量逻辑资源;最后通过实验验证, 所提出的FPGA光流计算方法的精度高于运行在PC平台的L&K方法, 卷积计算资源消耗明显降低。设计的系统可以完成1 280×1 024 pixel、60 frame/s输入视频的计算, 满足光流计算的实时性要求。

Optical flow estimation theory has been proposed in the field of machine vision for many years, but the high computational complexity of the algorithm limits its application in the industrial field. To meet the real-time requirements, a method was realized by FPGA in this study. To improve the accuracy of the algorithm and reduce the consumption of hardware resources, the Lucas and Kanade optical flow calculation method was improved. First, a two-level optical flow computation framework was designed to improve the accuracy. To address the problem of insufficient read-write rate of the external memory, it was proposed that when the image was read, it was sampled and cached to two spaces at the same time for subsequent computation. Considering the iterative correlation between the two levels of optical flow, we designed the data readout order to be stored in external memory. Then, in this study, a new convolution weight function was designed to reduce the volume of convolution, which was reduced by 73%, thus saving a lot of hardware resources. Experimental results indicate that the accuracy of the hardware implementation is higher than that of the Lucas and Kanade method on a PC, and the convolution computation is significantly reduced. This system fulfills the specified real time constrains of 60 images per second with 1 280×1 024 image resolution.