Adaptive Uncertainty-aware Pose Estimation for Intelligent Vehicles Based on Multi-Source Information

Accurate and robust pose estimation is a prerequisite for autonomous driving but remains challenging in complex urban environments due to sensor degradation, such as Global Navigation Satellite System (GNSS) blockage, Light Detection and Ranging (LiDAR) geometric degeneracy, and wheel slip. Furthermore, conventional fusion frameworks typically employ fixed noise parameters, failing to adapt to time-varying sensor reliability. To address these issues, this paper proposes a tightly-coupled adaptive pose estimation framework integrating GNSS, LiDAR, Inertial Measurement Unit (IMU), and vehicle chassis sensors via Factor Graph Optimization (FGO). First, a novel Two Degrees of Freedom (2-DoF) vehicle dynamics pre-integration factor is constructed using wheel speed and steering angle measurements, constraining lateral and heading drift during external perception outages. Second, a hierarchical confidence assessment mechanism is designed. This mechanism quantifies heterogeneous sensor reliability in real-time and dynamically adjusts fusion weights through a designed noise mapping model combined with a progressive recovery strategy. Extensive experiments on a real-world vehicle in complex scenarios and the M2DGR dataset demonstrate the system’s effectiveness. Compared to the standard LiDAR-Inertial Odometry (LIO) baseline, the proposed method reduces the translation and rotation Root Mean Square Error (RMSE) by 15.1% and 20.4%, respectively, while maintaining a real-time latency of 30–55 ms.