YOLO Acceleration using FPGA Architecture

Abstract

In recent years, convolutional neural networks (CNN) have achieved breakthrough developments. However, with the spatial and time complexity of CNN gradually increasing, it becomes a great challenge to implement larger CNN for mobile applications. Accelerated platforms based on FPGAs are gradually being studied due to its advantages such as high performance, low power consumption, reconfigurability, etc. In this paper, we implement the YOLO platform based on the Zynq board (which is released by Xillinx) and optimize its architecture combined the FPGA feature. The use case of pedestrian and cars recognition is demonstrated in real time, which outperforms the traditional CPU-based YOLO networks.