Abstract
Accurate vehicle localization is critical for the safe operation of autonomous vehicles. However, in GPS-denied environments such as tunnels or urban canyons, localization accuracy significantly deteriorates due to poor GPS signal quality. Traditional methods, such as extended Kalman filtering (EKF) and map matching, face limitations: EKF struggles with system nonlinearity and noise model assumptions, while map matching relies heavily on high-definition (HD) maps and often suffers in complex environments. Reinforcement learning (RL)-based approaches have shown promise, but they have typically been validated only under normal GPS conditions. This study proposes an RL-based method to enhance vehicle localization in GPS-denied environments. When GPS signals are unavailable, the framework estimates vehicle positions through dead reckoning using CAN-based vehicle speed and inertial measurement unit (IMU) heading information. The reinforcement learning policy subsequently corrects these estimated positions to reduce deviations from the HD map reference line. Additionally, GPS signal quality is used to identify GPS-denied conditions without requiring additional environment-detection sensors. The proposed method was evaluated in MATLAB/Simulink using recorded vehicle sensor data from selected tunnel scenarios, including segments before, during, and after GNSS outages. The experimental results showed lower RMSE values than the evaluated DR + GPS and EKF baselines under the tested conditions. These findings indicate the potential of the proposed method for improving localization accuracy in GPS-degraded environments, while broader validation is required to establish its general applicability and robustness.