Online Worker Selection Towards High Quality Map Collection for Autonomous Driving.
Xiaofeng Cao,Yan Li,Jiarong Han,Peng Yang,Feng Lyu,Deke Guo,Xuemin (Sherman) Shen
DOI: https://doi.org/10.1109/globecom38437.2019.9013516
2019-01-01
Abstract:Vehicle-based crowdsourcing is expected to be an economic yet efficient solution to build and maintain an accurate, fine-grained, and up-to-date environment map (i.e., high-definition map) for autonomous vehicles, which is an essential building block for safe and intelligent autonomous driving. However, how to select crowdsourcing workers with performance maximization is prudent and quite challenging since vehicles are highly dynamic and have unpredictable routes. In this paper, we study the worker selection problem for crowdsourced on-route map collection where the trade-off between the real-time worker exploration and exploitation is the main focus. Specifically, by adopting the multi-armed bandit model, we formulate a cumulative platform utility maximization problem. To solve this problem, we propose an Online Worker Selection (OWS) scheme, to learn drivers' performance and make worker selection decisions in real time. Essentially, two key designs are integrated in OWS: 1) performance transfer. If a new driver joins the crowdsourcing, we will initialize the new driver's performance based on the knowledge transferred from the existing drivers' records; and 2) marginal utility. Particularly, we carefully incorporate the platform utility to embody the marginal effect, i.e., repeated coverage by multiple vehicles on a certain road will undermine the utility. Based on the real-world vehicular GPS trace, we conduct extensive trace-driven simulations, and results demonstrate that our scheme can effectively obtain high-quality environment map, with on average 40.5% crowdsourcing utility gain over other benchmark schemes.