Degeneracy-Resilient LiDAR Odometry via Reflectance-Derived Correspondences

The key challenge in LiDAR odometry is estimating motion in geometrically degenerate environments, where standard geometrybased feature alignment often fails. LiDAR reflectance offers complementary information: it can be rendered as an image, letting the LiDAR to act like an active camera sensor and to add constraints where geometry is weak. We use these images to detect repeatable keypoints, match them across sweeps, and lift their locations to 3D, creating sparse and reflectance-informed correspondences. Our method follows a standard LiDAR-Inertial Odometry (LIO) pipeline. An Error-State Kalman Filter (ESKF) provides high-rate motion estimates for scan deskewing and for initializing ICP. We fuse reflectance-derived constraints into scan-to-map registration with a joint objective that combines a sparse point-to-point term with point-to-plane residuals, stabilizing motion directions that are otherwise weakly observable. We also select both reflectance and geometric correspondences to specifically constrain these weak directions. Experiments in geometrically degenerate and GNSS-denied settings, and even in presence of highly spatially anisotropic LiDAR acquisitions, show that adding reflectance-derived correspondences reduces drift and guides convergence toward the true pose.