Degeneracy-Aware Intensity-Assisted LiDAR–Inertial Odometry with Adaptive Photometric Weighting
Abstract
LiDAR–inertial odometry (LIO) is accurate in structurally rich environments but can become weakly observable in tunnels, stairways, and open terrain. This study introduces a degeneracy-aware, intensity-assisted LIO method that uses LiDAR reflectivity as an internal sensing modality without requiring a camera. Raw returns are projected onto a normalized panoramic intensity image, and image patches are selected according to their ability to complement the uninformative directions identified from the geometric information matrix. Point-to-plane, photometric, and inertial residuals are then fused in an iterated extended Kalman filter. Unlike fixed-scale intensity fusion, the proposed strategy adjusts the photometric residual scale according to the number of weak geometric directions and smooths this scale temporally. Experiments on the Newer College, ENWIDE, and GEODE datasets show that adaptive scaling reduces translational ATE RMSE, defined as the sequence-level root-mean-square of pose-wise translational absolute pose errors, by 11–40% on four ablation sequences. On GEODE-Stairs, the method obtains an ATE RMSE of 0.26 m, which is 46.9% lower than that of COIN-LIO. It also achieves the lowest ATE RMSE on each of the four tunnel sequences, with values ranging from 0.30 to 0.33 m. Mean processing times range from 15 to 25 ms per scan, corresponding to an average throughput of 40.0–66.7 scans/s on the evaluated platform. These results indicate that degeneracy-conditioned intensity fusion improves LIO robustness while maintaining average throughput compatible with real-time operation.