A Bayesian Optimization Approach for Online System Configuration of Edge Video Analytics

Surveillance cameras have been extensively deployed in urban areas. The captured video data requires real-time analysis, e.g., target detection and classification, to complete specific tasks, such as traffic control and public safety monitoring. Thanks to the advance of edge computing technology, hardware-constrained cameras can offload computationally intensive video analysis tasks to edge servers, alleviating the prolonged latency issues associated with traditional cloud-based video analytics. Constrained by the wireless transmission capacity, it is necessary to dynamically configure the resolution of video frames and the communication bandwidth to balance the video analysis accuracy and transmission delay. In practice, a major design difficulty is the lack of explicit models that describe the delay and accuracy of edge inference as a function of the configuration choices. For this, we propose an online algorithm based on Bayesian Optimization (BO) to jointly optimize the resolution selection and bandwidth allocation of edge video analytics systems on-the-fly. Simulation results demonstrate that the proposed BO-based method has comparable average performance as the “black-box” deep reinforcement learning algorithm, while achieving greater temporal stability and much higher fairness among the users. The results have showcased the importance of model interpretability offered by the BO-based method in enhancing the performance of dynamic edge video analytics systems.