Integration of Scan-Matching-based LiDAR Odometry with INS and GNSS for UAVs in Unstructured Natural Environments

Abstract Uncrewed Aerial Vehicles (UAVs) offer a low-cost and flexible means of collecting river hydrometry data, but the quality of the resulting measurements depends on accurate georeferencing. In unstructured natural environments, such as rivers, this is challenging as vegetated riverbanks and tree-canopy occlusion can degrade Global Navigation Satellite System (GNSS) reception, which in turn degrades the vehicle's Inertial Navigation System (INS)/GNSS navigation solution, worsening its position, velocity, and attitude estimates. This paper proposes an INS/GNSS/Light Detection and Ranging (LiDAR) architecture based on an error state Kalman filter, in which double-differenced GNSS measurements are tightly coupled, while velocity and attitude from a lightweight, scan-matching-based Iterative Closest Point (ICP) LiDAR odometry are loosely coupled. This odometry avoids reliance on both large local maps, and discernible geometric features, since water absorption and foliage scattering produce sparse, feature-poor point clouds. The proposed method is evaluated on two real-world airborne datasets from natural environments and compared against standalone Post-Processed Kinematic positioning, a state-of-the-art factor-graph approach, and a filter-based feature-dependent method. The solution outperforms all tested approaches across position, velocity, and attitude. Relative to a conventional double-differenced INS/GNSS integration, the scan-based solution reaches centimeter-level position accuracy across all axes in the river environment (Root Mean Square Error (RMSE) under 0.07 m). LiDAR measurements also contribute substantially to attitude estimation, particularly to heading, reducing the yaw RMSE by 66% compared to INS/GNSS. Moreover, the entire architecture runs in real time on a standard laptop, remaining well within a per-epoch processing budget suitable for onboard deployment.