Edge Video Analytics with Adaptive Information Gathering: A Deep Reinforcement Learning Approach

• Published in: IEEE transactions on wireless communications Vol. 22; no. 9; p. 1
• Main Authors: Wang, Shuoyao, Bi, Suzhi, Zhang, Ying-Jun Angela
• Format: Journal Article
• Language: English
• Published: New York IEEE 01.09.2023; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Accuracy; Algorithms; Bandwidths; Channel allocation; Decision analysis; Decomposition; Deep learning; Deep Reinforcement Learning; Degradation; Edge computing; Markov processes; Markove Decision Process; Mathematical analysis; Mobile computing; Multi-access Edge Computing; Network latency; Optimization; Public safety; Real-time systems; Resource Allocation; Resource management; Servers; Streaming media; Tradeoffs; Video Analytics; Visual analytics
• ISSN: 1536-1276; 1558-2248
• DOI: 10.1109/TWC.2023.3237202

With growing popularity of enormous public safety and transportation infrastructure cameras, there are increasing demands for automatic mobile video analytics. The emerging multi-access edge computing (MEC) technology has been recently applied to improve the accuracy-latency tradeoff of mobile video analytics. In this paper, we study an MEC-enabled multi-device video analytics system and formulate the problem as a Markov decision process (MDP) to meet two practical challenges: i) the absence of ground truth in real-time and ii) the content-varying degradation-accuracy relation. In particular, we aim to design an online joint frame degradation and bandwidth allocation algorithm with the time-varying function and limited feedback from each device. Thanks to the MDP formulation and n -step return technique, the long-term goal offers adaptive information gathering and thus improves the average accuracy and latency. For sample efficiency, we decompose the MDP problem into discrete degradation adaptation subproblems and continuous bandwidth allocation subproblems. Based on the decomposition, we propose a deep reinforcement learning (DRL) based framework, referred to as DBAG, to solve the decomposed subproblems. DBAG integrates model-based optimization and model-free DRL to solve the MDP problem with a discrete-continuous hybrid action space. Under various network setups and public datasets, DBAG greatly improves the accuracy-latency tradeoff.