Online resolution adaptation and resource allocation for edge-assisted video analytics

Abstract

Real-time analytics on video data from mobile devices demands intensive computation resources for the application like traffic monitoring and anomaly detection. Leveraging edge computing allows for the offloading of computation-intensive tasks to nearby edge servers, alleviating the constraints on resource-limited end-devices and reducing the long latency incurred by transmitting data to the cloud. When offloading the video data from multiple end-devices to multiple edge servers, factors such as video resolution, server selection, the allocation of bandwidth and computing resources have a substantial impact on key metrics like detection accuracy and real-time task success rate. The time-varying channel state and the dynamics of the event sequence also play a crucial role in decision-making. Building upon this foundation, our focus is on proposing an online solution for resolution adaptation and resource allocation that maximizes average utility over the long term, striking a balance between video analytics accuracy and real-time performance. The problem is formulated within a Markov decision process (MDP) framework. In view of the continuous state and action spaces, we propose an online solution based on the asynchronous advantage actor–critic (A3C) learning method. Extensive simulation results show that our proposed A3C-based solution exhibits faster convergence and superior performances compared with a series of benchmark algorithms under various settings, and thus achieves the best tradeoff between analytics accuracy and task success rate.