Intensity-enhanced LiDAR-inertial odometry with gradient flow sampling for degraded environments

High-precision localization is a fundamental requirement for autonomous robot navigation. However, in challenging LiDAR-degraded environments, sparse geometric structures, insufficient effective features, and interference from overlapping redundant points and cluttered noise often cause existing methods to drift severely, making accurate localization difficult. To address this, we propose a gradient flow sampled and intensity-enhanced LiDAR-Inertial Odometry (LIO) framework that improves matching efficiency and localization accuracy under geometric degeneracy. First, we propose a gradient flow-based point cloud sampling method that computes the distribution of point cloud gradient flows based on the observability of point cloud hyperplanes, minimizing sampling to suppress redundancy, and followed by a geometric consistency verification to reject noisy measurements. Second, to improve registration accuracy and robustness, we introduce an intensity-geometry fused point-pair association strategy that rates scan correspondences via intensity Kullback-Leibler (KL) divergence and geometric similarity to select the best match, integrates it into the point-to-plane iterative Extended Kalman Filter (iEKF) framework. Then, a dynamic factor during pose estimation adaptively balances geometric and photometric residuals across environments. Finally, extensive experiments on the Newer College, ENWIDE, DiTer++, and GEODE datasets show that the proposed algorithm outperforms the intensity-enhanced LIO algorithms on most sequences, with a 22.98% improvement in real-time performance compared to the baseline.