Variational Bayesian Adaptive KalmanNet for Intelligent Vehicle Localization Using GNSS/IMU Integration

Abstract

Accurate and reliable localization is essential for the safe deployment of intelligent vehicles (IVs). Global navigation satellite system (GNSS)/inertial measurement unit (IMU) fusion based on Bayesian filtering remains the most practical solution due to its low cost and broad applicability. However, classical filtering approaches are highly vulnerable to modeling inaccuracies and nonstationary disturbances, as motion models derived from low-grade IMUs are inevitably imperfect and measurement noise statistics vary significantly across driving conditions. This article presents a Variational Bayesian Adaptive KalmanNet (VBA-KalmanNet), an online model-correction learning framework for GNSS/IMU localization under simultaneous motion-model variation and time-varying GNSS noise. In the prediction stage, a recurrent neural network (RNN) learns corrections to the prior state estimate and its covariance, alleviating the effects of structural motion model mismatch. In the update stage, a variational Bayesian (VB) mechanism is employed to online infer the time-varying measurement noise covariance. Instead of learning a fixed gain or covariance mapping from training data, VBA-KalmanNet learns to correct the nominal IMU-driven prior online; the VB module then jointly infers the posterior state and the time-varying measurement noise covariance conditioned on this corrected prior. This unified design enables the proposed filter to adapt to complex vehicle dynamics and evolving noise characteristics without retraining. Experimental results on real-world datasets and field tests demonstrate that VBA-KalmanNet achieves superior localization accuracy and robustness under challenging and dynamic environments.