边缘视频处理的细粒度划分与重组部署算法

doi:10.12263/DZXB.20200767

摘要/Abstract

摘要：

随着视频数据的迅速增长，大规模视频处理业务需求急剧增加.如何及时处理视频数据获取有效信息，进而向用户快速提供视频分析业务是亟待解决的重要问题.针对此问题，提出一种面向大规模视频处理的边缘功能模块化及重组部署方法（EFMR）.该方法将视频处理业务下沉到网络边缘，利用网络功能虚拟化，将边缘服务器中的视频业务请求根据其内在相关性进行功能细粒度划分，按需匹配并最大化复用资源，实现重组部署，从而以较小代价实现边缘视频业务处理功能的平滑扩展.实验结果表明，EFMR方法不仅降低了边缘服务器的接入与响应时延、业务的推理时间，而且还节省了大量的计算资源，提高了视频处理业务部署速度.

关键词: 移动边缘计算, 网络功能虚拟化, 模块化, 重组, 细粒度

Abstract:

With the rapid growth of video data, the demand for large-scale video processing tasks increases dramatically. How to process video data in time to obtain effective information and provide users with video analysis services quickly is an important issue to be solved. Aiming at this problem, a new deployment method of Edge Functions Modularized and Reorganized (EFMR) for large-scale video processing is proposed. This method sinks video processing services to the edge of the network. Using network function virtualization, video service requests sent to the edge server are divided fine-grainedly based on their inherent process correlation, and resources are matched and redeployed on demand based on the division results. In this way, we can smoothly expand the edge video service processing capabilities at a small cost. Experimental results show that EFMR method not only greatly reduces the edge server’s access and response delay, reduces the inference time, but also saves a lot of computing resources of edge servers and speeds up the deployment of video processing services.

Key words: mobile edge computing, network function virtualization, modular, reorganized, fine-grained

中图分类号:

TN919.8

覃剑, 石昌伟, 张媛, 贾云健, 胡浩星. 边缘视频处理的细粒度划分与重组部署算法[J]. 电子学报, 2021, 49(11): 2152-2159.

QIN Jian, SHI Chang-wei, ZHANG Yuan, JIA Yun-jian, HU Hao-xing. Fine-Grained Partitioning and Reorganization Deployment Strategy of Edge Video Processing[J]. Acta Electronica Sinica, 2021, 49(11): 2152-2159.

图/表 13

参考文献 14

1	AbdulsalamY, ShailendraS, ShamimH M, et al. IoT big data analytics for smart homes with fog and cloud computing[J]. Future Generation Computer Systems, 2019,(2): 563 - 573.
2	GuoT. Cloud-based or on-device: An empirical study of mobile deep inference[A]. IEEE International Conference on Cloud Engineering (IC2E) [C]. Orlando, USA: IEEE, 2018. 184 - 190.
3	ZhouZ, LiaoH, GuB, et al. Robust mobile crowd sensing when deep learning meets edge computing[J]. IEEE Network, 2018, 32(4): 54 - 60.
4	KangY, HauswaldJ, GaoC, et al. Neurosurgeon: collaborative intelligence between the cloud and mobile edge[J]. ACM SIGPLAN Notices, 2017, 52(4): 615 - 629.
5	LiH, ShouG, HuY, et al. Mobile edge computing: Progress and challenges[A]. IEEE International Conference on Mobile Cloud Computing, Services, and Engineering[C]. Oxford, UK: IEEE, 2016. 83 - 84.
6	AhmedA, AhmedE. A survey on mobile edge computing[A]. International Conference on Intelligent Systems and Control[C]. Coimbatore, India, 2016. 1 - 8.
7	HuangY, MaX, FanX, et al. When deep learning meets edge computing[A]. IEEE International Conference on Network Protocols[C]. Toronto, Canada: IEEE, 2017. 1 - 2.
8	HuangY, ZhuY, FanX, et al. Task scheduling with optimized transmission time in collaborative cloud-edge learning[A]. IEEE International Conference on Computer Communication and Networks[C]. Hangzhou, China: IEEE, 2018. 1 - 9.
9	LiE, ZhouZ, ChenX. Edge intelligence: On-demand deep learning model co-inference with device-edge synergy[A]. Proceedings of the Workshop on Mobile Edge Communications[C]. Budapest, Hungary: ACM, 2018.31 - 36.
10	WangZ, XueG, QianS, et al. CampEdge: Distributed computation offloading strategy under large-scale Ap-based edge computing system for IoT applications[J]. IEEE Internet of Things Journal, 2020:6733 - 6745.
11	CampoloC, IeraA, MolinaroA, RuggeriG. MEC support for 5g-v2x use cases through docker containers[A]. IEEE Wireless Communications and Networking Conference[C]. Marrakesh, Morocco: IEEE, 2019. 1 - 6.
12	GirshickR, DonahueJ, DarrellT, et al. Rich feature hierarchies for accurate object detection and semantic segmentation[A]. IEEE Conference on Computer Vision and Pattern Recognition[C]. Columbus, USA: IEEE, 2014. 580 - 587.
13	EvgenyA, DenisM, SergeN. Comparison of regularization methods for ImageNet classification with deep convolutional neural networks[J]. AASRI Procedia, 2014,(6): 89 - 94.
14	UijlingsJRR, van de SandeKEA, GeversT, et al. Selective search for object recognition[J]. International Journal of Computer Vision, 2013, 104(2): 154 - 171.

层类型	浮点运算量（FLOPs）
卷积层1（11×11， 4， 96， ReLU）	105M
卷积层2（5×5， 1， 256， ReLU）	223M
卷积层3（3×3， 1， 384， ReLU）	149M
卷积层4（3×3， 1， 384， ReLU）	112M
卷积层5（3×3， 1， 256， ReLU）	74M
全连接层1（4096， ReLU）	37M
全连接层2（4096， ReLU）	16M
全连接层3（1000）	4M

层类型	浮点运算量（FLOPs）
卷积层1（11×11， 4， 96， ReLU）	105M
卷积层2（5×5， 1， 256， ReLU）	223M
卷积层3（3×3， 1， 384， ReLU）	149M
卷积层4（3×3， 1， 384， ReLU）	112M
卷积层5（3×3， 1， 256， ReLU）	74M
全连接层1（4096， ReLU）	37M
全连接层2（4096， ReLU）	16M
全连接层3（1000）	4M

过程	AlexNet	VGG16
疑似出现物体框	12s	12s
检测出目标	15s	18s

过程	AlexNet	VGG16
疑似出现物体框	12s	12s
检测出目标	15s	18s

模块	第一次业务（分类）	再一次业务（检测）
AlexNet特征提取	2s	0s
分类模块	12s	/
检测模块	/	13s
总时间	14s	13s