Degeneracy-aware continuous-time LiDAR-inertial-wheel odometry using geometric planelet normal statistics

Abstract

LiDAR–inertial odometry performs well in structured environments, but weak or anisotropic geometry can make the estimate drift. Ground vehicles encounter this problem in corridors, parking areas, and workshops, where motion is mainly planar and one horizontal direction may be poorly constrained. This paper presents a degeneracy-aware continuous- time LiDAR–inertial–wheel odometry framework. It estimates directional support from local planelet normals and uses that in- formation to adjust the sensor weights. In contrast to approaches that infer degeneracy primarily from the local numerical con- ditioning of the registration or estimator, the proposed method evaluates vehicle-body-frame planelet-normal support before the joint state update. When wheel and steering measurements are available, the resulting directional score adjusts the corre- sponding chassis residual. When they are unavailable, it instead attenuates LiDAR information in the weakly supported direction. The geometric planelet front-end and directional analysis were evaluated with the CT-LI LiDAR–IMU configuration on HeLiPR, which has no wheel-encoder data. The complete adaptive LiDAR–IMU–wheel framework was evaluated in three self-collected indoor scenes. Relative to FAST-LIO2, the LiDAR–IMU variant reduced APE RMSE by 33.0%, 41.6%, and 25.4% on the HeLiPR Town, Roundabout, and Bridge sequences, and the complete framework reduced APE RMSE by 47.5%, 25.9%, and 57.8% in the mixed corridor, underground parking, and workshop scenes. The indoor per-pose APE dispersion remained at or below 0.13 m. Processing one LiDAR frame took 83 ms at a 10 Hz input rate. Across the tested scenes, the method yields lower APE RMSE and error dispersion