Geometric Consistency–Guided LiDAR Odometry with Adaptive Feature Weighting in Unstructured Outdoor Environments

Geometric Consistency–Guided LiDAR Odometry

Abstract

Reliable 3D LiDAR odometry in complex, naturally unstructured, and perceptually challenging environments remains a critical problem for field robots. This paper presents a geometric-consistency-guided LiDAR odometry framework that assesses feature reliability before optimization and uses the resulting information to weight constraints during incremental pose estimation. First, a distance-adaptive neighborhood selection strategy compensates for range-dependent variations in point density. Local geometric analysis then classifies the feature points as stable or unstable according to their geometric consistency. Finally, constraints from stable and unstable features are assigned different weights within a nonlinear least-squares formulation. The framework is evaluated on self-collected agricultural datasets and public datasets covering botanical gardens, university environments in the GRACO dataset, and mines and caves in the DARPA Subterranean dataset. Evaluations using absolute trajectory error, relative pose error, ablation studies, and processing-frequency analyses at the component and system levels demonstrate competitive accuracy, reduced drift in challenging sequences, and computational performance that keeps pace with the input LiDAR stream. Overall, the results confirm the effectiveness of the proposed feature-reliability strategy for LiDAR odometry in diverse and challenging environments.