Bounded Planes and Wall Corners With Hierarchical Association for LiDAR-Inertial Odometry

Abstract

Planes, as higher-level geometric features, are widely employed to constrain motion, particularly in challenging scenes where point-based features degenerate. However, most existing techniques rely on modeling infinite planes, which often lack sufficient discriminability, are sensitive to pose drift, and can lead to ambiguous mapping. This method is intended for LiDAR-inertial odometry and mapping in structured indoor environments that approximately follow the Manhattan-world assumption. In this letter, planes are modeled as bounded patches using convex hulls to increase their discriminability. Furthermore, we extract wall corners—formed by three approximately mutually orthogonal structural planes (two vertical wall planes and one horizontal floor or ceiling plane) as distinctive structural features in indoor environments. These wall corners are parameterized as 6-DoF primitives to provide complementary translational and rotational constraints in the LIO system. A hierarchical mechanism is introduced to establish reliable data associations between planes and wall corners, enabling these features to serve as long-term landmarks for correcting LiDAR pose estimation. Experiments on self-collected and public datasets show that the proposed method achieves lower pose errors than the evaluated baselines.