Bayesian Online Learning for Sensory Data Analysis in Internet of Vehicles

By building a road information sharing system at the edge, namely EdgeSharing, the traffic safety problems caused by the perception blind spots of self-driving vehicles can be effectively alleviated. EdgeSharing is inherently a crowdsourcing system where sensory data are firstly uploaded from vehicles and then analyzed at the edge to maintain information updating. As such, uploading latency and analysis accuracy are two imperative components to measure the performance of EdgeSharing system, which however challenges the traditional Internet of Vehicles (IoV) that have no consideration for the diverse quality requirements of the content of sensory data. In this paper, we take video captured by on-board camera as an example and devise an adaptive configuration strategy for effectively data uploading and analysis at the edge. Naturally, configuration parameters affect system performance in an unknown and time-varying fashion because of the varying video content and network conditions. To address that, we adopt a Bayesian online learning method that learns the optimal configuration timely by observing historical system performance and real-time environmental information. Our strategy is able to maximize the data analysis accuracy at the edge under the constraint of desired frame rate, which makes EdgeSharing system accommodate complex vehicular environments. Simulation results indicate that compared to other schemes, our proposal finds the optimal configuration fastest and results in about 50% regret reduction.