Azimuth-LIO: Robust LiDAR-Inertial Odometry Via Azimuth-Aware Voxelization and Probabilistic Fusion
Author Names and Affiliations Omitted for Anonymous Review
AI summary
Problem
Voxel-based LiDAR-inertial odometry suffers from geometric inconsistencies when single-Gaussian models indiscriminately merge observations from conflicting viewpoints, leading to biased covariance estimation and degraded localization accuracy in complex or occluded environments.
Approach
The method decomposes each voxel into azimuth-sectorized substructures modeled by anisotropic Gaussians, applies a direction-weighted registration metric to prioritize sensor-aligned views, and updates the map using a Bayesian fusion framework that adapts confidence weights based on viewing angle.
Key results
- Decomposes voxels into azimuth-partitioned substructures with directionally weighted registration
- Introduces a Bayesian fusion framework for perspective-consistent map updates
- Achieves superior accuracy and robustness on M2DGR, MCD, and SubT-MRS benchmarks
- Maintains real-time computational efficiency comparable to existing voxel-based methods
Why it matters
Provides a robust, uncertainty-aware solution for autonomous navigation in complex, occluded, or dynamic environments where traditional single-Gaussian voxel models fail.
Abstract
Voxel-based LiDAR–inertial odometry (LIO) is ac- curate and efficient but can suffer from geometric inconsistencies when single-Gaussian voxel models indiscriminately merge obser- vations from conflicting viewpoints. To address this limitation, we propose Azimuth-LIO, a robust voxel-based LIO framework that leverages azimuth-aware voxelization and probabilistic fusion. Instead of using a single distribution per voxel, we discretize each voxel into azimuth-sectorized substructures, each modeled by an anisotropic 3D Gaussian to preserve viewpoint-specific spatial features and uncertainties. We further introduce a direction- weighted distribution-to-distribution registration metric to adap- tively quantify the contributions of different azimuth sectors, followed by a Bayesian fusion framework that exploits these confidence weights to ensure azimuth-consistent map updates. The performance and efficiency of the proposed method are evaluated on public benchmarks including the M2DGR, MCD, and SubT-MRS datasets, demonstrating superior accuracy and robustness compared to existing voxel-based algorithms.