Improving the robustness of vehicle localization by remodeling the position error distribution in urban environments

Abstract

The integration of global navigation satellite system (GNSS) and inertial navigation system (INS) provides the fundamental location-based service for autonomous driving and intelligent transportation systems. However, traditional Gaussian assumption-based integration methods cannot provide stable localization performance in complex urban environments. Therefore, the GNSS position error distribution is remodeled in this study to enhance the robustness of vehicle localization. Firstly, the distribution characteristics of GNSS positioning errors across multiple scenarios are systematically analyzed, based on which a Gaussian–Student’s t mixture distribution is used to refine the error representation and construct a hierarchical linear Gaussian state-space model. Subsequently, a low-complexity adaptive optimization mechanism is designed to fine-calibrate the prior measurement stochastic model. Finally, joint iterative estimation of state parameters and auxiliary variables is conducted through the variational Bayesian framework. This method ensures rapid responsiveness to environmental changes while achieving a synergistic optimization between adaptability and robustness. Field experiments conducted in typical urban environments demonstrate that the proposed algorithm can achieve superior overall performance, better than traditional robust methods with improvements of 38.1% and 28.8% in position and velocity respectively. Especially in representative challenging scenarios, i.e. urban canyon and bridge underpass, the proposed method achieves position RMSE improvements of over 22.6% and 46.1% compared with traditional approaches, respectively. It highlights the strong adaptability and robustness in complex urban environments.