DRL-Based User Scheduling for Collaborative Localization in GNSS Denied Environment

Accurate vehicle state estimation is crucial for autonomous driving, yet it is challenging in a global navigation satellite system (GNSS) denied environment. Inertial navigation system (INS) equipped in each vehicle can achieve qualified short-term localization but the performance gets degraded during long-term operation due to cumulative errors. In this paper, we leverage the relative position observed by on-board sensors to achieve centralized collaborative localization. In particular, we focus on uplink user scheduling to choose the most urgent vehicle to transfer its INS and millimeter-wave radar data to a roadside unit (RSU). Considering a situation with limited spectrum, we utilize a deep reinforcement learning (DRL) algorithm to solve the problem of user scheduling, as it enables the optimization of scheduling strategies through continuous learning in dynamics vehicular environments. The agent, i.e., the RSU, chooses an appropriate V2I link to occupy the spectrum to receive data from the user at each time step, and employs Kalman filter and extended Kalman filter algorithms to predict and update vehicle states. Simulation results validate the superiority of the proposed scheme in reducing the average root mean square error.