Real Time Object Detection Based on FPGA with Big Data
Peiyue Zhang, Zhan Xu, Pengcheng Liu, Yiyang Zhao, Lian Wang, Yuntao Ma, Jian Wang · 2018
Over recent years, various types of vision-based object detection models are proposed for avoiding traffic accidents or self-driving. However, in vehicle, the electronic control of driving part is achieved by SOC(System On Chip). Due to limitations on data capacity and compute power of SOC, the existing methods cannot be directly implemented to vehicles. In this paper, we present Simple-FBY(Simple-FPGA Based YOLO), a vehicle and pedestrian detection model based on FPGA which is a real time, high speed, low power computing platform. It can provide accuracy, speed and be capable of learning. To this end, we introduce a specifically designed training data set, which consists of pictures of other public self-driving data sets and a Beijing's streets data set collecting by ourselves. Compared with state-of-the-art model YOLO, our FPGA-based model can provide high performance per watt of power consumption. We also adjust the YOLO algorithm to make it more adaptable to SOC. In the experiment, we compare performance of Simple-FBY on GPU and Xilinx Zynq-Z1 FPGA Board. As a result, our model achieves better accuracy on vehicle and pedestrian detection and shows a good speed performance at FPGA platform.