1.ShanghaiTech;2.Shanghai Institute of Technical Physics, Chinese Academy of Sciences
基于深度学习的目标检测算法取得了很大成功，显著超越了传统算法，在很多场景下甚至可以和人类相媲美。不同于可见光相机，红外相机可以在黑暗环境下识别物体，可以用于安防和无人驾驶等领域。本文提出了面向嵌入式设备的轻量级目标检测算法，并采用赛灵思的Ultrascale+MPSoC ZU3EG FPGA加速并部署该算法。加速器运行在350MHz的时钟频率下吞吐量达到了551FPS，功耗仅有8.4W。在准确率方面，该算法在FLIR数据集下IoU指标达到了73.6%。在性能方面，相比于之前相同逻辑资源下性能最好的硬件加速器Ultranet，该加速器设计将吞吐量提高了2.59倍，功耗降低了2.04倍。
Object detection algorithm based on deep learning has achieved great success, significantly better than the effect of traditional algorithms, and even surpassed human in many scenarios. Unlike RGB cameras, infrared cameras can see objects even in the dark, which can be used in many fields like surveillance and autonomous driving. In this paper, a lightweight target detection algorithm for embedded devices is proposed, and the algorithm is accelerated and deployed using Xilinx’s Ultrascale+MPSoC ZU3EG FPGA. The accelerator runs at a 350MHz frequency clock with throughput of 551FPS and power of only 8.4W. As for accuracy, the intersection over union (IoU) of the algorithm achieves an accuracy of 73.6% on FILR datasets. Compared to the previous work, the accelerator design improves performance by 2.59× and reduces power consumption by 2.04×.