Real-Time Localization for Autonomous Vehicles via Fusion of Low-Rate LiDAR Poses and GNSS/IMU Data

Robust real-time localization is a prerequisite for safe autonomous driving. While LiDAR-based methods provide precise local pose estimation, their performance often degrades in environments lacking distinct features like tunnels or open highways or when the pre-built map is outdated due to environmental changes. In contrast, GNSS ensures global consistency but suffers from signal instability in urban canyons. To address these limitations, this paper proposes a robust localization framework based on parallel dual Error-State Kalman Filters (ESKFs). The system maintains two independent estimators: LiDAR-IMU and GNSS-IMU fusion. A key contribution is the dynamic covariance scaling mechanism that translates NDT matching quality into probabilistic uncertainty, allowing the system to assess LiDAR reliability in real-time. By comparing the posterior covariance of the two filters, the framework dynamically selects the more trustworthy estimate and maintains localization continuity under practically common complementary GNSS/LiDAR degradation. Extensive evaluations were conducted on both high-speed custom datasets (up to 110 km/h) and the KITTI residential benchmark. On degraded KITTI maps, the proposed method achieved RMSEs of 0.6727 m and 0.5497 m on Sequences 0027 and 0034, respectively, while preserving a 200 Hz output stream through IMU-rate state propagation.