A Scalable MDP-Based Sensing and Processing Framework for Vehicular Networks

Abstract

Smart Vehicles (SV) generally have on-board processing, wireless communication and sensing capabilities. They are useful for vehicular applications like autonomous driving (AD) or in-vehicle augmented reality (AR). One way to enhance the performance of these applications is by enabling SVs to exchange messages and sensor readings with other nearby SVs. However, due to the mobility and resource limitation of SVs, the availability of sensing and processing resources of nearby SVs and the communication links between SVs (V2V) are unreliable. Thus, it is a challenge to develop efficient sensing and processing schemes for SVs. We first propose a Markov decision processes (MDP) based sensing and proposing framework for SVs. The framework models the uncertainties and can be solved to obtain an optimum sensing and processing policy. Second, we point out that the spatial information required to obtain sensing policy, needs an extremely big state space for representation, making this framework unscalable. Thirdly, we propose separating the sensor selection problem from the sensing and processing scheme and explain how this simplifies the problem and makes it scalable. We then propose a maximum flow minimum cost based sensor selection heuristic. Finally, we compare the performance obtained by applying our heuristic with that of our original MDP based scheme. The results show that our heuristics performance is nearly as good as the original scheme while also increasing scalability.