AD-LIO: Adaptive Voxel LiDAR–Inertial Odometry Based on Scene Degradation Factor

Accurate and stable self-localization across the full domain scenarios is critical for autonomous driving and the Internet of Vehicle (IoV). When autonomous vehicles enter degraded scenarios such as tunnels, self-localization may fail due to the sparse geometric features and high similarity structures. This paper introduces AD-LIO, an adaptive voxel LiDAR inertial odometry framework designed for IoVs, which can quantitatively evaluate the degradation of scenarios to guide the framework in dealing with challenging scenarios. Points-based registration is utilized to guarantee the full details of the feature-lack scene can be described. To ensure this robust registration, a novel point-to-plane strategy, which employs the LiDAR geometry and intensity constraints, is incorporated into the point-to-plane matching. The framework can adaptively change voxel size to adjust the scenario condition in realtime by proposing a lightweight scene degradation factor. This factor can be simply calculated and attached to the scan-to-map optimization process. Extensive experiments under degradation and non-degradation scenarios show that the proposed AD-LIO framework outperforms state-of-the-art methods in robustness and estimation precision. The framework also achieves a favorable balance between precision and computational efficiency across LiDAR configurations from 16 to 128 beams. Furthermore, AD-LIO enables reliable large-scale pose estimation over trajectories exceeding 60 km. Our demonstration video is available now at https://youtu.be/Qy2mD4347mQ.