面向边缘实时视频分析的资源高效持续学习框架

doi:10.16182/j.issn1004731x.joss.25-0590

系统仿真学报 ›› 2026, Vol. 38 ›› Issue (2): 294-306.doi: 10.16182/j.issn1004731x.joss.25-0590

• 学习与集成框架 • 上一篇

面向边缘实时视频分析的资源高效持续学习框架

吴舒霞¹^,²^,³, 张俊杰¹^,²^,³, 陈德珑¹^,²^,³, 陈哲毅¹^,²^,³

^1.福州大学计算机与大数据学院，福建福州 350116
^2.大数据智能教育部工程研究中心，福建福州 350002
^3.福建省网络计算与智能信息处理重点实验室(福州大学)，福建福州 350116

收稿日期:2025-06-24 修回日期:2025-09-11 出版日期:2026-02-18 发布日期:2026-02-11
通讯作者: 陈哲毅
第一作者简介:吴舒霞(2001-)，女，硕士生，研究方向为边缘计算、实时视频分析、资源优化。
基金资助:
国家自然科学基金(62202103);福建省杰出青年科学基金(2025J010020);中央引导地方科技发展资金(2022L3004);福建省科技经济融合服务平台(2023XRH001);福厦泉国家自主创新示范区协同创新平台(2022FX5)

Resource-efficient Continuous Learning Framework for Edge Real-time Video Analytics

Wu Shuxia¹^,²^,³, Zhang Junjie¹^,²^,³, Chen Delong¹^,²^,³, Chen Zheyi¹^,²^,³

^1.College of Computer and Data Science, Fuzhou University, Fuzhou 350116, China
^2.Engineering Research Center of Big Data Intelligence, Ministry of Education, Fuzhou 350002, China
^3.Fujian Key Laboratory of Network Computing and Intelligent Information Processing (Fuzhou University), Fuzhou 350116, China

Received:2025-06-24 Revised:2025-09-11 Online:2026-02-18 Published:2026-02-11
Contact: Chen Zheyi

摘要/Abstract

摘要：

通过在网络边缘部署轻量化模型，边缘系统可提供实时视频分析服务。由于模型训练与实际部署之间的差异会导致数据漂移，为构造与真实环境相匹配的轻量化模型，提出了一种面向边缘实时视频分析的资源高效持续学习框架(continuous learning framework for edge real-time video analytics，CL4VA)。引入了一种面向感兴趣区域粒度的精度下降预测器以高效选取实时视频流中的关键样本；构建了一种双层混合样本池以自适应触发模型持续学习并避免灾难性遗忘问题；设计了一种基于DRL的控制器以决定完成模型重训练的合适时机。仿真结果表明：CL4VA相较于基准方法可降低平均8.65%的延迟和提升最高5.57%的精度。同时，CL4VA的核心组件仅需极低的在线开销。

关键词: 边缘计算, 实时视频分析, 资源优化, 持续学习, 深度强化学习

Abstract:

By deploying lightweight models at the network edge, edge systems can provide services of real-time video analytics. However, due to the data drift caused by the discrepancy between model training and actual deployment, it is challenging to construct lightweight models that match real-world environments. To address this challenge, a resource-efficient continuous learning framework for edge real-time video analytics (CL4VA) was proposed. A region of interest-granularity predictor for accuracy degradation was introduced to efficiently select key samples from real-time video streams.A two-layer mixed sample pool was constructed to adaptively trigger the model's continuous learning and avoid the issue of catastrophic forgetting.A DRL-based controller was designed to determine the appropriate time to complete model re-training. The simulation results show that compared to the baseline method, CL4VA can reduce the average latency by 8.65% and increase the accuracy by up to 5.57%. Moreover, the core components of CL4VA require extremely low online overhead.

Key words: edge computing, real-time video analytics, resource optimization, continuous learning, DRL

中图分类号:

TP393

吴舒霞,张俊杰,陈德珑等 . 面向边缘实时视频分析的资源高效持续学习框架[J]. 系统仿真学报, 2026, 38(2): 294-306.

Wu Shuxia,Zhang Junjie,Chen Delong,et al . Resource-efficient Continuous Learning Framework for Edge Real-time Video Analytics[J]. Journal of System Simulation, 2026, 38(2): 294-306.

图/表 9

图1

表1

图2

表2

图3

图4

图5

表3

图6

参考文献 28

[1]	Xu Renjie, Razavi Saiedeh, Zheng Rong. Edge Video Analytics: A Survey on Applications, Systems and Enabling Techniques[J]. IEEE Communications Surveys & Tutorials, 2023, 25(4): 2951-2982.
[2]	袁牧, 张兰, 姚云昊, 等. 面向智能物联网的资源高效模型推理综述[J]. 计算机学报, 2024, 47(10): 2247-2273.
	Yuan Mu, Zhang Lan, Yao Yunhao, et al. Resource-efficient Model Inference for AIoT: A Survey[J]. Chinese Journal of Computers, 2024, 47(10): 2247-2273.
[3]	李超, 李贾宝, 丁才昌, 等. 基于DRL的边缘监控任务卸载与资源分配算法[J]. 系统仿真学报, 2024, 36(9): 2113-2126.
	Li Chao, Li Jiabao, Ding Caichang, et al. Edge Surveillance Task Offloading and Resource Allocation Algorithm Based on DRL[J]. Journal of System Simulation, 2024, 36(9): 2113-2126.
[4]	Gao Xitong, Zhao Yiren, Dudziak Łukasz, et al. Dynamic Channel Pruning: Feature Boosting and Suppression[EB/OL]. (2019-01-28) [2025-06-12]. .
[5]	Banitalebi-Dehkordi Amin, Vedula Naveen, Pei Jian, et al. Auto-split: A General Framework of Collaborative Edge-cloud AI[C]//Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining. New York: Association for Computing Machinery, 2021: 2543-2553.
[6]	Huang Kai, Gao Wei. Real-time Neural Network Inference on Extremely Weak Devices: Agile Offloading with Explainable AI[C]//Proceedings of the 28th Annual International Conference on Mobile Computing and Networking. New York: Association for Computing Machinery, 2022: 200-213.
[7]	Chen Gang, He Shengyu, Meng Haitao, et al. PhoneBit: Efficient GPU-accelerated Binary Neural Network Inference Engine for Mobile Phones[C]//2020 Design, Automation & Test in Europe Conference & Exhibition (DATE). Piscataway: IEEE, 2020: 786-791.
[8]	Chen Yuhan, Ravindranath L, Deng Shuo, et al. Glimpse: Continuous, Real-time Object Recognition on Mobile Devices[C]//Proceedings of the 13th ACM Conference on Embedded Networked Sensor Systems. New York: Association for Computing Machinery, 2015: 155-168.
[9]	Yuan Mu, Zhang Lan, He Fengxiang, et al. InFi: End-to-end Learnable Input Filter for Resource-efficient Mobile-centric Inference[C]//Proceedings of the 28th Annual International Conference on Mobile Computing And Networking. New York: Association for Computing Machinery, 2022: 228-241.
[10]	Li Yuanqi, Padmanabhan A, Zhao Pengzhan, et al. Reducto: On-camera Filtering for Resource-efficient Real-time Video Analytics[C]//Proceedings of the Annual conference of the ACM Special Interest Group on Data Communication on the applications, technologies, architectures, and protocols for computer communication. New York: Association for Computing Machinery, 2020: 359-376.
[11]	Ling Neiwen, Wang Kai, He Yuze, et al. RT-mDL: Supporting Real-time Mixed Deep Learning Tasks on Edge Platforms[C]//Proceedings of the 19th ACM Conference on Embedded Networked Sensor Systems. New York: Association for Computing Machinery, 2021: 1-14.
[12]	Jiang Shiqi, Lin Zhiqi, Li Yuanchun, et al. Flexible High-resolution Object Detection on Edge Devices with Tunable Latency[C]//Proceedings of the 27th Annual International Conference on Mobile Computing and Networking. New York: Association for Computing Machinery, 2021: 559-572.
[13]	Yang Zheng, Wang Xu, Wu Jiahang, et al. EdgeDuet: Tiling Small Object Detection for Edge Assisted Autonomous Mobile Vision[J]. IEEE/ACM Transactions on Networking, 2023, 31(4): 1765-1778.
[14]	Mullapudi R T, Chen S, Zhang Keyi, et al. Online Model Distillation for Efficient Video Inference[C]//2019 IEEE/CVF International Conference on Computer Vision (ICCV). Piscataway: IEEE, 2019: 3572-3581.
[15]	Khani M, Ananthanarayanan G, Hsieh K, et al. RECL: Responsive Resource-efficient Continuous Learning for Video Analytics[C]//20th USENIX Symposium on Networked Systems Design and Implementation (NSDI 23). Boston: USENIX Association, 2023: 917-932.
[16]	Zhang Jingwei, Cheng Long, Liu Cong, et al. Cost-aware Scheduling Systems for Real-time Workflows in Cloud: An Approach Based on Genetic Algorithm and Deep Reinforcement Learning[J]. Expert Systems with Applications, 2023, 234: 120972.
[17]	Cai Huaiguang, Zhou Zhi, Huang Qianyi. Online Resource Allocation for Edge Intelligence with Colocated Model Retraining and Inference[C]//IEEE INFOCOM 2024 - IEEE Conference on Computer Communications. Piscataway: IEEE, 2024: 1900-1909.
[18]	Kong Yuxin, Yang Peng, Cheng Yan. Edge-assisted On-device Model Update for Video Analytics in Adverse Environments[C]//Proceedings of the 31st ACM International Conference on Multimedia. New York: Association for Computing Machinery, 2023: 9051-9060.
[19]	Shubha S S, Shen Haiying. AdaInf: Data Drift Adaptive Scheduling for Accurate and SLO-guaranteed Multiple-model Inference Serving at Edge Servers[C]//Proceedings of the ACM SIGCOMM 2023 Conference. New York: Association for Computing Machinery, 2023: 473-485.
[20]	Zhang Ziyang, Zhao Yang, Chang Mingching, et al . E 4: Energy-efficient DNN Inference for Edge Video Analytics Via Early Exiting and DVFS[C]//Proceedings of the Thirty-ninth AAAI Conference on Artificial Intelligence and Thirty-seventh Conference on Innovative Applications of Artificial Intelligence and Fifteenth Symposium on Educational Advances in Artificial Intelligence. Palo Alto: AAAI Press, 2025: 1165-1173.
[21]	Wang Liang, Qu Xiaoyang, Wang Jianzong, et al. Gecko: Resource-efficient and Accurate Queries in Real-time Video Streams at the Edge[C]//IEEE INFOCOM 2024 - IEEE Conference on Computer Communications. Piscataway: IEEE, 2024: 481-490.
[22]	Bhardwaj R, Xia Zhengxu, Ananthanarayanan G, et al. Ekya: Continuous Learning of Video Analytics Models on Edge Compute Servers[C]//19th USENIX Symposium on Networked Systems Design and Implementation (NSDI 22). Renton: USENIX Association, 2022: 119-135.
[23]	Schulman J, Wolski F, Dhariwal P, et al. Proximal Policy Optimization Algorithms[EB/OL]. (2017-08-28) [2025-06-12]. .
[24]	Zeng Xiao, Fang Biyi, Shen Haichen, et al. Distream: Scaling Live Video Analytics with Workload-adaptive Distributed Edge Intelligence[C]//Proceedings of the 18th Conference on Embedded Networked Sensor Systems. New York: Association for Computing Machinery, 2020: 409-421.
[25]	Huang Wenhui, Zhou Yanxin, He Xiangkun, et al. Goal-guided Transformer-enabled Reinforcement Learning for Efficient Autonomous Navigation[J]. IEEE Transactions on Intelligent Transportation Systems, 2024, 25(2): 1832-1845.
[26]	Leal-Taixé Laura, Milan A, Reid I, et al. MOTChallenge 2015: Towards a Benchmark for Multi-Target Tracking[EB/OL]. (2015-04-08) [2025-06-12]. .
[27]	Yu F, Chen Haofeng, Wang Xin, et al. BDD100K: A Diverse Driving Dataset for Heterogeneous Multitask Learning[C]//2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2020: 2633-2642.
[28]	Wen Longyin, Du Dawei, Cai Zhaowei, et al. UA-DETRAC: A New Benchmark and Protocol for Multi-object Detection and Tracking[J]. Computer Vision and Image Understanding, 2020, 193: 102907.

特征	含义	作用
area	候选RoI面积与整帧图像面积之比	刻画了RoI在图像中的占比，当RoI较小或被部分遮挡时，其在低分辨率场景下容易丢失
ratio	候选RoI的外接矩形宽高比	描述了RoI的形态特征，反映了遮挡场景带来的比例异常
circularity	RoI中所有前景轮廓平均圆形度的反值	反映了RoI轮廓的几何规则，在光照不均或遮挡导致轮廓断裂时，该值将升高
convexity	RoI中所有前景轮廓的平均凸度	反映了RoI轮廓的凹陷特征，在遮挡场景下RoI轮廓会出现凹陷，导致该值下降
position	候选RoI中心点相对整帧图像中心的归一化位置	表示RoI在图像中的分布情况，位于边缘位置的 RoI 更容易受到拍摄角度、光照不均或遮挡的影响而难以被检测到

数据集	指标	Vanilla	Ekya	AdaInf	CL4VA	Oracle
UA-DETRAC	延迟/ms	19.98	34.48	35.65	30.23	41.42
UA-DETRAC	mAp/%	20.20	58.32	58.44	63.89	70.15
BDD100K	延迟/ms	21.78	38.07	42.18	33.39	46.45
BDD100K	mAp/%	28.68	32.46	33.48	35.89	36.50
MOT	延迟/ms	16.45	28.83	31.24	29.42	39.20
MOT	mAp/%	73.69	78.23	79.98	80.64	78.93
Real-Camera	延迟/ms	17.12	30.74	30.61	27.51	39.89
Real-Camera	mAp/%	59.14	94.66	97.25	97.75	100

面向边缘实时视频分析的资源高效持续学习框架

Resource-efficient Continuous Learning Framework for Edge Real-time Video Analytics

RichHTML

PDF (PC)

可视化

摘要/Abstract

引用本文

使用本文

图/表 9

参考文献 28

相关文章 15

编辑推荐

Metrics

本文评价

[1]	江明, 何韬. 基于深度强化学习的带容量约束车辆路径问题求解[J]. 系统仿真学报, 2025, 37(9): 2177-2187.
[2]	倪培龙, 毛鹏军, 王宁, 杨孟杰. 基于改进A-DDQN算法的机器人路径规划[J]. 系统仿真学报, 2025, 37(9): 2420-2430.
[3]	陈真, 吴卓屹, 张霖. 深度强化学习中策略表征研究简述[J]. 系统仿真学报, 2025, 37(7): 1753-1769.
[4]	伍国华, 曾家恒, 王得志, 郑龙, 邹伟. 基于深度强化学习的四旋翼航迹跟踪控制方法[J]. 系统仿真学报, 2025, 37(5): 1169-1187.
[5]	张森, 代强强. 改进型深度确定性策略梯度的无人机路径规划[J]. 系统仿真学报, 2025, 37(4): 875-881.
[6]	李敏, 张森, 曾祥光, 王刚, 张童伟, 谢地杰, 任文哲, 张滔. 基于深度强化学习的四足机器人单腿越障轨迹规划[J]. 系统仿真学报, 2025, 37(4): 895-909.
[7]	王贺, 许佳宁, 闫广宇. 基于深度强化学习的AGV行人避让策略研究[J]. 系统仿真学报, 2025, 37(3): 595-606.
[8]	张斌, 雷永林, 李群, 高远, 陈永, 朱佳俊, 鲍琛龙. 基于强化学习的导弹突防决策建模研究[J]. 系统仿真学报, 2025, 37(3): 763-774.
[9]	黄思进, 文佳, 陈哲毅. 面向边缘车联网系统的智能服务迁移方法[J]. 系统仿真学报, 2025, 37(2): 379-391.
[10]	费帅迪, 蔡长龙, 刘飞, 陈明晖, 刘晓明. 舰船防空反导的目标分配方法研究[J]. 系统仿真学报, 2025, 37(2): 508-516.
[11]	白臻祖, 侯一帜, 何章鸣, 魏居辉, 周海银, 王炯琦. 考虑随机扰动的动态武器目标分配优化[J]. 系统仿真学报, 2025, 37(12): 2967-2980.
[12]	郑家瑜, 麦著学, 陈哲毅. 数字孪生云边网络下服务缓存与计算卸载优化[J]. 系统仿真学报, 2025, 37(11): 2741-2753.
[13]	邸剑, 万雪, 姜丽梅. 基于精英指导和随机搜索的进化强化学习[J]. 系统仿真学报, 2025, 37(11): 2877-2887.
[14]	徐忠锴, 储晨阳, 解凯, 赵睿卓, 柯文俊. 基于SC-PPO的高比例新能源电力系统优化调度方法[J]. 系统仿真学报, 2025, 37(10): 2511-2521.
[15]	梁秀满, 刘子良, 刘振东. 基于深度强化学习的改进RRT算法路径规划[J]. 系统仿真学报, 2025, 37(10): 2578-2593.