Edge-Assisted Video Transmission with Adaptive Key Frame Selection: A Hierarchical DRL Approach

This paper proposes a novel hierarchical deep reinforcement learning (HDRL) framework for edge-assisted real-time video transmission in Industrial Internet of Things (IIoT). The system model consists of a gateway, multiple edge servers, and a central controller. The gateway performs key frames selection to compress video chunks from terminal cameras, which are then transmitted through multi-hop links to edge servers for video analysis. The central controller determines the key frames selection and routing path for each video chunk to minimize the transmission delay while ensuring video data accuracy. Different from the existing work, we investigate video transmission under system dynamics that the bandwidth of each link is time-varying and both real-time and accuracy requirements of each chunk are unpredictable. We decompose such problem into routing path and key frames selection sub-problems. To this end, we introduce a deep Q network-based optimal routing approach and an adaptive key frames selection approach to solve the two sub-problems, respectively. An HDRL training framework is further developed to integrate these two approaches jointly for improving the overall performance. Simulation results show the superiority of the proposed solution over counterparts.