摘要
当前交通、安防等领域广泛应用摄像头采集视频进行分析,传统将视频流直接上传到云平台处理的方式面临接入量受限、时延大等问题;边云协同架构下,将部分视频流卸载到边缘服务器可降低时延,可缓解云服务压力。考虑到视频分析任务对准确率、时延和能耗都有一定要求,提出通过同时控制视频帧的分辨率、边缘服务器部署卷积神经网络(Convolution Neural Network, CNN)模型的策略以及边云卸载决策,来最大化视频分析准确率,同时满足长期平均时延和能耗约束的问题。利用李雅普诺夫随机优化理论将原优化问题转化为每个时隙的独立优化问题,并采用蚁群优化算法求解得到动态卸载优化策略,包括视频帧的分辨率选择、边缘服务器部署哪些CNN模型以及边云卸载决策。仿真实验结果表明,所提动态卸载策略相比其他基线方案能够在满足约束的情况下获得更高的视频分析准确率。
Video analytics is widely adopted in fields of transportation and security nowadays.However,traditional method of offloading video streams direct to the cloud for processing suffers from problems such as restricted access and high latency.Thus,an edge-cloud collaborative computing architecture is proposed,where some of the video streams are offloaded to the edge server to reduce latency and alleviate the load of cloud.Considering the requirements for accuracy,latency,and energy consumption in video analytics tasks,a strate-gy that simultaneously controls the resolution of video frames,the convolution neural network(CNN)model deployment on the edge serv-er,and the edge-cloud offloading decisions is proposed to maximize video analytics accuracy while satisfying long-term average latency and energy consumption constraints.By utilizing the Lyapunov stochastic optimization theory,the original optimization problem is trans-formed into a series of independent optimization problems in each time slot,which are solved by using ant colony algorithm to obtain dy-namic offloading strategy including the selection of video frame resolution,the CNN models to be deployed on edge servers,and the edge-cloud offloading decisions.Simulation results show that the proposed dynamic offloading strategy achieves higher video analytics accuracy compared to other baseline schemes while satisfying the constraints.
作者
童佳慧
李越
李燕君
毛科技
TONG Jiahui;LI Yue;LI Yanjun;MAO Keji(College of Computer Science and Technology,Zhejiang University of Technology,Hangzhou Zhejiang 310023,china)
出处
《传感技术学报》
北大核心
2025年第1期128-134,共7页
Chinese Journal of Sensors and Actuators
基金
浙江省自然科学基金重点项目(LZ25F020009)
国家自然科学基金项目(61772472)。
关键词
边云协同计算
卸载决策
李雅普诺夫理论
蚁群优化算法
视频分析
edge-cloud collaborative computing
offloading decision
Lyapunov theory
ant colony optimization algorithm
video analytics
