通信学报 ›› 2020, Vol. 41 ›› Issue (10): 109-115.doi: 10.11959/j.issn.1000-436x.2020192

所属专题: 边缘计算

• 学术论文 • 上一篇    下一篇

面向视频监控基于联邦学习的智能边缘计算技术

赵羽,杨洁,刘淼,孙金龙,桂冠   

  1. 南京邮电大学通信与信息工程学院,江苏 南京 210003
  • 修回日期:2020-07-07 出版日期:2020-10-25 发布日期:2020-11-05
  • 作者简介:赵羽(1996- ),男,江苏泰州人,南京邮电大学博士生,主要研究方向为基于深度学习的智能视频分析|杨洁(1980- ),女,江苏南京人,博士,南京邮电大学讲师,主要研究方向为基于深度学习的智能视频分析|刘淼(1988- ),男,江苏淮安人,博士,南京邮电大学讲师,主要研究方向为基于深度学习的智能通信|孙金龙(1988- ),男,河南洛阳人,博士,南京邮电大学讲师,主要研究方向为基于深度学习的智能通信|桂冠(1982- ),男,安徽枞阳人,博士,南京邮电大学教授,主要研究方向为基于深度学习的物理层无线通信技术
  • 基金资助:
    工信部重大专项基金资助项目(TC190A3WZ-2);国家自然科学基金资助项目(61901228)

Federated learning based intelligent edge computing technique for video surveillance

Yu ZHAO,Jie YANG,Miao LIU,Jinlong SUN,Guan GUI   

  1. College of Telecommunications and Information Engineering,Nanjing University of Posts and Telecommunications,Nanjing 210003,China
  • Revised:2020-07-07 Online:2020-10-25 Published:2020-11-05
  • Supported by:
    The Major Project of the Ministry of Industry and Information Technology of China(TC190A3WZ-2);The National Natural Science Foundation of China(61901228)

摘要:

随着全球数据量的激增,集中式云计算无法提供低时延、高效率的视频监控服务。基于此,提出分布式边缘计算模型,在边缘端直接处理视频数据,减少网络的传输压力,缓解中央云服务器的计算负担,降低视频监控系统的处理时延。结合联邦学习算法,采用轻量级神经网络,分场景训练模型,并将其部署于计算能力受限的边缘设备上。实验结果表明,对比通用神经网络模型,所提方法检测准确度提高18%,模型训练时间有效减少。

关键词: 联邦学习, 深度学习, 边缘计算, 轻量级神经网络, 目标检测

Abstract:

With the explosion of global data,centralized cloud computing cannot provide low-latency,high-efficiency video surveillance services.A distributed edge computing model was proposed,which directly processed video data at the edge node to reduce the transmission pressure of the network,eased the computational burden of the central cloud server,and reduced the processing delay of the video surveillance system.Combined with the federated learning algorithm,a lightweight neural network was used,which trained in different scenarios and deployed on edge devices with limited computing power.Experimental results show that,compared with the general neural network model,the detection accuracy of the proposed method is improved by 18%,and the model training time is reduced.

Key words: federated learning, deep learning, edge computing, lightweight neural network, object detection

中图分类号: 

No Suggested Reading articles found!