Lightweight High-Bandwidth LiDAR-Inertial Odometry Employing Sparse Incremental Hash-Voxels

Abstract

This work introduces a lightweight LIO framework employing incremental voxels for enhanced efficiency. We leverage a sparse voxel data structure, replacing the tree structure in the Point-LIO open-source framework. Through hash table-managed voxel indexes, we achieve rapid K nearest neighbor search within nearly one voxel size with constant complexity query speed. This approach significantly reduces the time cost associated with tree nodes construction, balancing, and iteration compared to the tree-like structures. Experimental results demonstrate that our proposed enhancement achieves an average speed-up of 21.5% compared to Point-LIO in publicly available datasets. Moreover, it reduces drift by an average of approximately 20(m) and ATE by 1.41(m) under sparse point cloud input conditions.