ISTD-LIOM: Direct LiDAR-inertial odometry and mapping with intensity-enhanced stable triangle descriptor

Abstract

To address the cumulative drift problem of light detection and ranging (LiDAR)-inertial odometry (LIO) in long-duration localization and mapping tasks, this paper proposes a LiDAR-inertial odometry and mapping system, intensity-enhanced stable triangle descriptor-LiDAR-inertial odometry and mapping (ISTD-LIOM), based on the intensity-enhanced stable triangle descriptor (ISTD). This system, built on the FAST-LIO2 front-end architecture, achieves global consistency localization through loop closure detection and global optimization. First, we design the ISTD descriptor by combining geometric descriptors of triangles (including vertex plane normal vectors and edge lengths) with local intensity distribution descriptors to form a compact, rotation-invariant feature representation. Next, an adaptive keyframe management mechanism is constructed, which filters keyframes based on inter-frame relative poses and generates a descriptor database. A hybrid retrieval strategy is then proposed, which combines descriptor similarity matching and spatial distance filtering, forming an efficient loop closure candidate recognition mechanism. After applying plane iterative closest point (ICP) refinement and geometric-intensity consistency validation, the loop closure constraints are integrated into a pose graph optimization framework, correcting odometry drift. Experiments on the KITTI dataset demonstrate that the ISTD-LIOM system significantly enhances map global consistency while maintaining real-time computational performance.