摘要
低空航拍视角往往背景复杂、目标小而多,难以检测识别,实际飞行器又有低延时和高处理速度的要求。针对该应用场景,文中提出完整的嵌入式FPGA解决方案,多种技术模块化协同地实现深度学习应用的高效部署。由飞行器定向采集制作数据集,基于SSD优化深度学习网络框架;通过高层次综合的方式设计辅助计算核,并定制DPU加速核,共同组成硬件平台,由Vitis工具链统合编译生成嵌入式操作系统镜像;基于Vitis AI技术生成网络推断函数库,基于OpenCL技术设计高层次综合硬件调度函数库,以动态链接库的方式兼容于基于Python设计的多线程主机应用程序中。测试结果表明,测试集上的均值平均精度(mAP)为0.55,实时处理速度约为20 f/s。文中方案在指标上满足了应用需求,并且可推广至其他深度学习的嵌入式部署设计。
The low-altitude aerial photography perspective often has complex background,small and many targets,which are difficult to detect and identify.Actually,there are the requirements of low delay and high processing speed for aircrafts.For this application scenario,a complete solution based on embedded FPGA is proposed,in which multiple technologies are modularized and coordinated to achieve the efficient deployment of deep learning applications.The aircraft is used to directionally collect and produce dataset,and the framework of deep learning network is optimized based on SSD.The auxiliary computing core is designed by means of the high-level synthesis method,and the DPU acceleration core is customized to form a hardware platform,which is integrated and compiled by the Vitis tool chain to generate a mirror image of the embedded operating system.A network inference function library is generated based on Vitis AI technology,and the high-level comprehensive hardware scheduling function library is designed based on OpenCL technology,which are compatible with multi-threaded host applications designed on the basis of Python in the way of dynamic link library.The testing results show that the mAP on the test set is 0.55,and the real-time processing speed is about 20 f/s.The solution can meet the application requirements in terms of indicators and can be extended to other embedded deployment designs of deep learning.
作者
吴李煜
张紫龙
张华君
田野
常胜
WU Liyu;ZHANG Zilong;ZHANG Huajun;TIAN Ye;CHANG Sheng(School of Physics and Technology,Wuhan University,Wuhan 430072,China;Hubei Aerospace Vehicle Research Institute,Wuhan 430040,China)
出处
《现代电子技术》
2022年第2期1-6,共6页
Modern Electronics Technique
基金
国家自然科学基金项目(61874079)
武汉市应用基础前沿项目(2018010401011289)。
关键词
目标检测
嵌入式FPGA
深度学习
高层次综合
硬件加速
精确检测
target detection
embedded FPGA
deep learning
high-level synthesis
hardware acceleration
precision detection
作者简介
吴李煜(1994-),男,江苏无锡人,硕士,研究方向为嵌入式系统;张紫龙(1986-),男,四川雅安人,高级工程师,研究方向为智能微型飞行器;张华君(1988-),男,湖北十堰人,工程师,研究方向为人工智能嵌入式应用;田野(1996-),男,湖北枝江人,硕士,研究方向为深度学习算法;常胜(1980-),男,湖北武汉人,博士,教授,研究方向为新型电子器件与电路。
