基于FPGA+ARM异构平台的道路交通目标检测设计  

Design of Road Traffic Target Detection System Based on FPGA+ARMHeterogeneous Platform

作　　者：胡胜 王逸风 易文涛 贺岚晴 宋海娜 刘聪 HU Sheng;WANG Yifeng;YI Wentao;HE Lanqing;SONG Haina;LIU Cong(School of Electrical and Electronic Engineering,Hubei University of Technology,Wuhan 430068,China;Hubei Provincial Key Laboratory of Solar Energy Efficient Utilization and Energy Storage Operation Control,Hubei University of Technology,Wuhan 430068,China)

机构地区：[1]湖北工业大学电气与电子工程学院,湖北武汉430068 [2]湖北工业大学太阳能高效利用及储能运行控制湖北省重点实验室,湖北武汉430068

出　　处：《无线电通信技术》2024年第5期1008-1015,共8页Radio Communications Technology

基　　金：国家自然科学基金(62202148)。

摘　　要：针对当前目标检测技术在智能汽车上所需的体积小、能效比高、检测速度快、精度高等要求,提出了一种新型嵌入式道路交通目标检测设计。该方案以YOLOv3为基础网络模型,通过新增网络检测层和网络剪枝技术分别提升网络的检测精度和检测速度,在硬件端以深度处理单元(Deep Learning Processing Unit,DPU)为核心搭建了底层硬件平台,对网络的卷积计算进行并行加速,改进后的模型通过量化和编译后可以部署至FPGA+ARM异构平台。经测试,在KITTI数据集上的检测精度为85.32%,功耗为8.2 W,检测帧率为31.2 Hz,算力功耗比达到58.3 GOPs/W,是RTX 2060 super型GPU的3.9倍,Intel i7-12400型CPU的9.3倍。实验结果表明,该方案满足道路交通目标检测设计要求,相较常用目标检测平台GPU,所提方案的部署空间小、功耗低,更适于灵活部署在空间紧凑、能源供给受限的智能汽车中。In response to the requirement of small size,high energy efficiency,fast detection speed,and high accuracy for current object detection technology in intelligent vehicles,a novel embedded road traffic object detection design is proposed.This design uses YOLOv3 as the base network model and enhances detection accuracy and speed by adding new network detection layers and employing network pruning techniques.On the hardware side,a hardware platform is built around the Deep Learning Processing Unit(DPU)to parallelize the network’s convolutional computations.The improved model,after quantization and compilation,can be deployed on an FPGA+ARM heterogeneous platform.Testing on the KITTI dataset shows a detection accuracy of 85.32%,power consumption of 8.2 W,detection speed of 31.2 Hz,and a computational power efficiency of 58.3 GOPs/W,which is 3.9 times that of the RTX 2060 Super GPU and 9.3 times that of the Intel i7-12400 CPU.Experimental results indicate that this design meets the requirement for road traffic object detection.Compared to commonly used object detection platforms like GPUs,the proposed solution is more suitable for flexible deployment in intelligent vehicles with limited space and energy supply due to its smaller deployment footprint and lower power consumption.

关 键 词：智能汽车 目标检测 深度学习 嵌入式 FPGA 

分 类 号：U461.99[机械工程—车辆工程]