Scholarlink

<p>Traditional task scheduling and offloading approaches in edge-cloud environment often fail to achieve global optimality and struggle with NP-hard computational models involving hierarchical queues and limited resources. This paper uses the UCF-Crime dataset, which consists of actual surveillance video frames, to present an intelligent edge-cloud framework for Real-Time (RT) human activity identification. The system preprocesses the video frames, detects humans using YOLOv5, estimates the poses using YOLOv8, and extracts higher-level semantic features using EfficientNetB7. A hybrid Deep Learning (DL) model using Conv1D, BiLSTM, as well as attention mechanisms, have classified behaviors into normal or abnormal with good accuracy. We also propose a Dynamic-DQN-based scheduler that dynamically executes computational tasks between the edge node and cloud node, powered by a Q/Q/K queuing model, providing the necessary advanced QoS along with real-time system metrics. The proposed scheme achieves an accuracy of 98.21 and an AUC value of 98.60. Overall, the proposed scheme provides a scalable, energy-efficient, and rapid solution for smart monitoring applications.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

EdgeRL-surveillance: adaptive decision making for resource-constrained video analytics using reinforcement learning

  • 期刊论文

  • 发表日期：

    2026年7月20日

查看全文
  • Sandhya Rani Nallola,
  • A. Vadivel

摘要

Traditional task scheduling and offloading approaches in edge-cloud environment often fail to achieve global optimality and struggle with NP-hard computational models involving hierarchical queues and limited resources. This paper uses the UCF-Crime dataset, which consists of actual surveillance video frames, to present an intelligent edge-cloud framework for Real-Time (RT) human activity identification. The system preprocesses the video frames, detects humans using YOLOv5, estimates the poses using YOLOv8, and extracts higher-level semantic features using EfficientNetB7. A hybrid Deep Learning (DL) model using Conv1D, BiLSTM, as well as attention mechanisms, have classified behaviors into normal or abnormal with good accuracy. We also propose a Dynamic-DQN-based scheduler that dynamically executes computational tasks between the edge node and cloud node, powered by a Q/Q/K queuing model, providing the necessary advanced QoS along with real-time system metrics. The proposed scheme achieves an accuracy of 98.21 and an AUC value of 98.60. Overall, the proposed scheme provides a scalable, energy-efficient, and rapid solution for smart monitoring applications.