Adaptive Point and Directional Constraint Sampling for Robust LiDAR Odometry in Narrow and Elongated Environments

Abstract

LiDAR-based simultaneous localization and mapping (SLAM) achieves high localization and mapping accuracy in narrow geometric environments. However, existing research primarily focuses on multisensor fusion and pose estimation, with insufficient attention given to generic downsampling operations. Such downsampling may discard observation points that significantly contribute to pose estimation and implicitly alter the point cloud sampling distribution. We propose a unified point sampling framework for elongated environments, consisting of preestimation sampling prior to pose estimation and in-optimization resampling during pose optimization. First, an elongated-aware sampling (EAS) method is proposed prior to pose estimation, which performs point cloud segmentation and adaptive voxelization using a lagged sliding window to enable geometry-aware sampling. Planar and edge features are selected based on curvature, with dominant planes down-sampled and salient edge points retained. Subsequently, a directionwise degeneracy-graded point resampling (DGPR) method is proposed during pose estimation, which evaluates localizability by jointly considering information distribution and absolute information content. Directionwise localizability is assessed using sliding-window-based adaptive thresholds, and informative measurements are resampled along degenerate directions. Extensive experiments on representative elongated datasets demonstrate that the proposed sampling strategy consistently improves localization accuracy and robustness compared with state-of-the-art LiDAR odometry (LO) and LiDAR-inertial odometry (LIO) methods.