This work proposes an incremental Generalized Iterative Closest Point (GICP) based tightly-coupled LiDAR-inertial odometry (LIO), iG-LIO, which addresses the challenges of integrating GICP into real-time LIO. The main contributions are as follows.

• The raw GICP constraints are tightly-coupled with IMU constraints in a Maximum A Posteriori (MAP) estimation.
• A voxel-based surface covariance estimator (VSCE) is proposed to improve the efficiency and accuracy of the surface covariance estimation. Compared to the kd-tree based methods, VSCE reduces processing time in dense scans while maintaining the accuracy of iG-LIO in sparse and small FOV scans.
• An incremental voxel map is designed to represent the probabilistic models of surrounding environments. Compared to non-incremental methods (e.g., DLIO), it successfully reduces the time cost required for the nearest neighbor search and map management.
• Extensive datasets collected from different FOV LiDARs are adopted to evaluate the efficiency and accuracy of the proposed iG-LIO. Even though iG-LIO keeps identical parameters across all datasets, the results show that it is more efficient than Faster-LIO and achieves competitive performance compared to state-of-the-art LIO systems.

The experiment video can be found on YouTube , bilibili.

The paper is available in PDF.

The docker-based standard development environment is available at https://github.com/zijiechenrobotics/ig_lio_workspace

1️⃣ Ubuntu and ROS

Ubuntu >= 18.04. And Ubuntu 20.04 is recommended.

2️⃣ GCC & G++ (only for Ubuntu 18.04)

gcc & g++ >= 9

3️⃣ TBB (only for Ubuntu 18.04)

TBB >= 2020. Please follow https://github.com/oneapi-src/oneTBB

4️⃣ livox_ros_driver

git clone https://github.com/Livox-SDK/Livox-SDK
cd Livox-SDK
mkdir build
cd build
cmake ..
make -j
sudo make install

5️⃣ glog

sudo apt-get install -y libgoogle-glog-dev

cd <your workspace>
mkdir src
cd src
git clone https://github.com/zijiechenrobotics/ig_lio_workspace.git
git clone https://github.com/Livox-SDK/livox_ros_driver
cd ..
catkin_make

Download NCLT from http://robots.engin.umich.edu/nclt/

source devel/setup.bash
roslaunch ig_lio lio_nclt.launch

Download NCD from https://ori-drs.github.io/newer-college-dataset/

source devel/setup.bash
roslaunch ig_lio lio_ncd.launch

Download ULHK from https://github.com/weisongwen/UrbanLoco

source devel/setup.bash
roslaunch ig_lio lio_ulhk.launch

Download AVIA from https://drive.google.com/drive/folders/1CGYEJ9-wWjr8INyan6q1BZz_5VtGB-fP (fast-lio) and https://github.com/ziv-lin/r3live_dataset (r3live)

source devel/setup.bash
roslaunch ig_lio lio_avia.launch

The fast-lio datasets miss the gravitational constant in the accelerometer. Please edit the avia.ymal

# for fast-lio
enable_acc_correct: true

# for r3live
enable_acc_correct: false

Download Botanic Garden from https://github.com/robot-pesg/BotanicGarden

source devel/setup.bash
# for avia
roslaunch ig_lio lio_bg_avia.launch
# for velodyne
roslaunch ig_lio lio_bg_velodyne.launch

1️⃣ Edit .yaml files in ig_lio/config

• lidar_topic: LiDAR topic name.
• imu_topic: IMU topic name.
• lidar_type: The type of LiDAR you use. Only support for Velodyne, Ouster, and Livox.
• min_radius & max_radius: A range filter to remove laser point from the robot itself.
• enable_ahrs_initalization: Set true or false. If the IMU message has orientation channel, iG-LIO can be initialized via AHRS.
• enable_acc_correct: Set true or false. If the accelerometer miss the gravitational constant, please set true (e.g., fast-lio2 datasets).
• gravity: Make sure the gravity is correct. Some datasets are 9.81, some datasets are -9.81, and even zero (e.g., ULHK). A simple debugging method is to observe the glog message. The normal range of ba_norm is 0~0.5.
• t_imu_lidar & R_imu_lidar: The extrinsic parameters from LiDAR frame to IMU frame (i.e. the IMU is the base frame).

2️⃣ Launch iG-LIO

source devel/setup.bash
roslaunch ig_lio <your launch file name>.launch

rosbay play <your rosbag>

We use abbreviations for all sequences due to limited space. The full names of all sequences are presented below.

We aligned the mapping results of iG-LIO with Google Earth and found that iG-LIO retains global consistency maps.

Thanks for citing iG-LIO (RA-L 2024) if you use any of this code.

# IEEE Robotics and Automation Letters ( Early Access )
@ARTICLE{10380742,
  author={Chen, Zijie and Xu, Yong and Yuan, Shenghai and Xie, Lihua},
  journal={IEEE Robotics and Automation Letters}, 
  title={iG-LIO: An Incremental GICP-based Tightly-coupled LiDAR-inertial Odometry}, 
  year={2024},
  volume={},
  number={},
  pages={1-8},
  doi={10.1109/LRA.2024.3349915}}

Thanks for the below great open-source project for providing references to this work.

1. LOAM (J. Zhang and S. Singh. LOAM: Lidar Odometry and Mapping in Real-time)
2. FAST-LIO
3. Faster-LIO
4. LINS
5. SLICT

Thanks for the following public dataset.

1. NCLT
2. Newer College Dataset
3. Botanic Garden
4. R3live

What can’t iG-LIO do?

• The extremely narrow environment (e.g., some sequence in the Hilti SLAM Challenge).
• The scene is very open and devoid of geometric features.