R-VoxelMap: Accurate Voxel Mapping with Recursive Plane Fitting for Online LiDAR Odometry

Haobo Xi, Shiyong Zhang*, Qianli Dong, Yunze Tong, Songyang Wu, Jing Yuan, Xuebo Zhang · 2026

Read R-VoxelMap: Accurate Voxel Mapping with Recursive Plane Fitting for Online LiDAR Odometry by Haobo Xi, Shiyong Zhang*, Qianli Dong, Yunze Tong, Songyang Wu, Jing Yuan, Xuebo Zhang (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-12377.txt.

Introduction

Haobo Xi and six co authors work on how a moving LiDAR unit keeps track of where it is. Their tool builds maps from small cubes called voxels. Good maps help the unit place itself with care as it moves. Past tools in this line fit planes with all points in a cube at once.

That all in approach can go wrong in a few ways. stray points can pull a plane off course. Large planes can be split when they should stay whole. Planes from distinct surfaces can also be merged by mistake. The team puts forward R-VoxelMap, which fits planes step by step in a recursive way. The aim is a more exact map and sharper placement.

Catalog record

CREATORHaobo Xi, Shiyong Zhang*, Qianli Dong, Yunze Tong, Songyang Wu, Jing Yuan, Xuebo Zhang
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE40 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2601-12377

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

R-VoxelMap: Accurate Voxel Mapping with Recursive Plane Fitting for Online LiDAR Odometry Haobo Xi, Shiyong Zhang*, Qianli Dong, Yunze Tong, Songyang Wu, Jing Yuan, Xuebo Zhang Abstract This paper proposes R-VoxelMap, a novel voxel mapping method that constructs accurate voxel maps using a geometry-driven recursive plane fitting strategy to enhance the localization accuracy of online LiDAR odometry. VoxelMap and its variants typically fit and check planes using all points in a voxel, which may lead to plane parameter deviation caused by outliers, over segmentation of large planes, and incorrect merging across different physical planes. To address these issues, R-VoxelMap utilizes a geometry-driven recursive construction strategy based on an outlier detect-and-reuse pipeline. Specifically, for each voxel, accurate planes are first fitted while separating outliers using random sample consensus (RANSAC). The remaining outliers are then propagated to deeper octree levels for recursive processing, ensuring a detailed representation of the environment. In addition, a point distribution-based validity check algorithm is devised to prevent erroneous plane merging. Extensive experiments on diverse open-source LiDAR(-inertial) simultaneous localization and mapping (SLAM) datasets validate that our method achieves higher accuracy than other state-of-the-art approaches, with comparable efficiency and memory usage. Code will be available on GitHub. 1 1 1 https://github.com/NKU-MobFly-Robotics/R-VoxelMap I Introduction The rapid development of 3D light detection and ranging (LiDAR) technology has greatly accelerated the adoption of LiDAR(-inertial) odometry across diverse applications. Due to its high accuracy, this technology has been widely utilized and extensively researched over the past decade. To support accurate and real-time state estimation, mainstream LiDAR odometry approaches increasingly rely on map structures that can accurately represent the environment while supporting efficient maintenance and registration. As the most classic map structure, k-d tree enables point cloud registration via k-nearest neighbor (KNN) search and is widely used in LiDAR(-inertial) odometry [ 21 , 10 , 8 , 15 ] . Fast-LIO2 [ 14 ] proposes an incremental k-d tree, enabling more efficient map updates. However, such tree-based methods still face time-consuming KNN search and complex tree updates. To further improve efficiency, voxel-based methods have been proposed and become the mainstream [ 1 , 19 , 17 ] . These methods partition space into voxels and apply spatial hashing to enable fast point matching and registration, significantly enhancing efficiency while reducing the implementation complexity of map maintenance. Fig. 1: VoxelMap plane features suffers from outlier sensitivity (c, yellow-marked planes are erroneously shifted due to surrounding outliers), over-segmentation (d, white-boxed planar features are overly segmented, resulting in large uncertainty), and incorrect plane merging (d, red-marked regions show erroneous plane merging, generating large planar structures in areas that are not physically planar), all of which reduce localization accuracy. Our method effectively addresses these limitations. As one of the most widely adopted voxel-based map structures, VoxelMap [ 19 ] enables odometry methods to achieve improved accuracy and efficiency using probabilistic plane features. However, since it performs plane fitting using all points within a voxel, it inevitably suffers from two major limitations. First, the fitted planes are sensitive to outliers, even a small number of outliers can significantly distort the estimated plane parameters, thereby degrading localization and mapping accuracy (as shown in Fig. 1 (c)). While stricter plane calculation thresholds help reduce outlier inclusion, they often lead to over-segmentation (as shown in Fig. 1 (d)), resulting in fragmented planes with fewer points and slower uncertainty convergence, which is not conducive to accurate pose estimation. To alleviate this issue, VoxelMap’s variants [ 13 , 16 ] further merge coplanar planes to reconstruct larger planes with more point support and lower parameter uncertainty. However, since plane fitting before merging still relies on all points within a voxel, the initially extracted small planes may exhibit parameter deviations, which can eventually result in incomplete merging or incorrect merging. Second, points from different physical surfaces may be erroneously merged into the same plane, causing map distortion (as shown in Fig. 1 (d)). To address these challenges, we propose R-VoxelMap: a geometry-driven recursive plane fitting strategy for accurate voxel mapping. Different from other variants, we do not follow a split-and-merge strategy. Instead, we directly use the proposed outlier detect-and-reuse pipeline during the coarse-to-fine plane feature extraction, which effectively suppresses the influence of outliers. By simultaneously ensuring accurate plane feature extracting and preserving multi-scale geometric information, the proposed method effectively enhances odometry accuracy. Our contributions can be concluded as follows: 1. A novel plane feature voxel map structure and construction strategy, R-VoxelMap, which is constructed via geometry-driven recursive plane fitting. By performing an outlier detect-and-reuse pipeline in each recursive iteration, R-VoxelMap effectively suppresses the influence of outliers, enables more accurate map representation and reduces plane over-segmentation. 2. A point distribution-based plane validity check algorithm that projects and clusters point cloud on the RANSAC-fitted plane to effectively prevent the incorrect merging of different physical planes. 3. Extensive comparative experiments on multiple open-source LiDAR(-inertial) SLAM datasets demonstrate that our method achieves higher state estimation accuracy than other state-of-the-art approaches. The code will be open-sourced to benefit the community. Fig. 2: System overview of the LiDAR-based odometry using R-VoxelMap. The red-highlighted modules represent the core of R-VoxelMap, a geometry-driven recursive map construction algorithm based on an outlier detect-and-reuse pipeline. The blue modules are specifically designed to support the fundamental functions of R-VoxelMap. II Related Work The map structure for LiDAR odometry is required to support accurate and efficient point cloud registration, and can be categorized into tree-based and voxel-based approaches. Based on these map structures, LiDAR odometry can perform accurate point cloud registration using iterative closest points (ICP) [ 2 ] or generalized ICP (GICP) [ 9 ] algorithms. Classic LiDAR odometry methods [ 21 , 11 ] and a range of LiDAR-inertial odometry systems [ 10 , 8 , 15 ] , organize plane and edge feature points using k-d tree and perform registration via ICP. The efficiency of these methods is often constrained by the slowness of k-d tree searches and frequent tree reconstructions. Fast-LIO2 [ 14 ] uses a novel incremental k-d tree to reduce the overhead of frequent reconstructions. But its performance is still constrained by KNN search and rebalancing overhead. Faster-LIO [ 1 ] replaces the tree-based map structures with incremental voxels, which improves computational efficiency while maintaining localization accuracy. However, these point cloud-based mapping algorithms typically treat matched points as accurate representations of the environment, without accounting for the uncertainty introduced by sensor measurements and pose estimation. VoxelMap [ 19 ] employs an adaptive-size, coarse-to-fine voxel-based map structure based on hash map and octrees, where each voxel contains a probabilistic plane feature. This structure supports accurate modeling of the environment while considering uncertainty. Current VoxelMap variants typically improve odometry accuracy by merging coplanar plane features. For example, VoxelMap++ [ 13 ] reduces plane uncertainty and sensitivity to noise by merging potentially coplanar planes using a union-find algorithm. C3P-VoxelMap [ 16 ] , on the other hand, proposes an on-demand voxel merging strategy based on locality-sensitive hashing. These works confirm that merging coplanar points to form more robust plane features enhances pose estimation accuracy. However, all these VoxelMap-based methods [ 19 , 13 , 16 ] fit plane features using all points within each voxel, making them sensitive to outliers, which can distort the

Read the complete text