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

Introduction

R-VoxelMap is a voxel mapping method that improves localization accuracy in online LiDAR odometry by using a geometry-driven recursive plane fitting strategy. Our code is based on VoxelMap and primarily addresses the issues where VoxelMap and its variants typically fit and check planes using all points in a voxel, leading to parameter deviation due to outliers, over-segmentation of large planes, and incorrect merging across different physical planes. The main changes are as follows:

1. R-VoxelMap performs an outlier detect-and-reuse pipeline in each recursive iteration, effectively suppressing outlier influence, improving map accuracy, and reducing plane over-segmentation.

2. R-VoxelMap uses a point distribution-based plane validity check strategy, projecting and clustering point clouds on the RANSAC-fitted plane to prevent incorrect merging of different physical planes.

Related paper

Related paper available on arxiv and IEEE RAL 2026.

The framework of Lidar(-inertial) odometry based on R-VoxelMap.

The construction steps of R-VoxelMap. (outlier detect-and-reuse pipeline)

1. Dependencies

The required dependencies are same as VoxelMap.

1.1. PCL && Eigen

PCL>= 1.8, Eigen>= 3.3.4

1.2. livox_ros_driver

Follow livox_ros_driver Installation.

2. Build

Clone the repository and catkin_make:

    cd ~/$A_ROS_DIR$/src
    git clone https://github.com/NKU-MobFly-Robotics/R-VoxelMap.git
    cd ..
    catkin_make
    source devel/setup.bash

• Remember to source the livox_ros_driver before build (follow 1.2 livox_ros_driver).

3. Run

3.1 Run on rosbag

KITTI odometry dataset for example. The KITTI rosbag we used can be downloaded from Baidu Netdisk.

    cd ~/$R_VOXEL_MAP_ROS_DIR$
    source devel/setup.bash
    roslaunch rvoxelmap kitti.launch
    rosbag play kitti_00.bag --delay 1

• Make sure the topics match the rostopics you are using before running.
• Make sure log file path is modified to a suitable path.
• It is not recommended to enable the visualization option for R-VoxelMap plane features, as it may affect efficiency. If visualization is needed, you can use rosservice /rvoxelmap/publish_all for one-time visualization.
• --delay 1 is used to ensure rosbag's first frame data can be captured by the algorithm.

3.2 Run for test

We provide programs for testing the construction and update of R-VoxelMap.

Test construction (using PCD file)

    cd ~/$R_VOXEL_MAP_ROS_DIR$
    source devel/setup.bash
    roslaunch rvoxelmap test.launch 

• Make sure test.launch's pcd_file parameter points to your own PCD file.

Test update (using rosbag)

    cd ~/$R_VOXEL_MAP_ROS_DIR$
    source devel/setup.bash
    roslaunch rvoxelmap test_update.launch 
    rosbag play test.bag

• You need to prepare a test bag file containing point cloud data in PointCloud2 format.
• Make sure test_update.launch input point cloud topic remap to the correct topic.
• It is also not recommended to enable the visualization option for R-VoxelMap plane features, as it may affect efficiency. If visualization is needed, you can use rosservice /publish_again or /publish_all for one-time visualization.

Acknowledgments

Our code is built on top of VoxelMap. We would like to acknowledge the contributions of the VoxelMap team for providing such a valuable framework.