iG-LIO

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.

1. Build

1.1 Docker Container

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

1.2 Build from source

1.2.1 Prerequisites

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

1.2.2 Build

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

2. Run

2.1 NCLT Dataset

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

source devel/setup.bash
roslaunch ig_lio lio_nclt.launch

2.2 NCD Dataset

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

source devel/setup.bash
roslaunch ig_lio lio_ncd.launch

2.3 ULHK Dataset

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

source devel/setup.bash
roslaunch ig_lio lio_ulhk.launch

2.4 AVIA Dataset

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

2.5 Botanic Garden Dataset

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

2.6 Run with your own dataset

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>

3. Details about all sequences in the paper

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

Abbreviation	Name	Distance(km)	Sensor Type
nclt_1	2012-01-15	7.58	Velodyne HDL-32E
nclt_2	2012-04-29	3.17	Velodyne HDL-32E
nclt_3	2012-05-11	6.12	Velodyne HDL-32E
nclt_4	2012-06-15	4.09	Velodyne HDL-32E
nclt_5	2013-01-10	1.14	Velodyne HDL-32E
ncd_1	01_short_experiment	1.61	Ouster OS1-64
ncd_2	02_long_experiment	3.06	Ouster OS1-64
ncd_3	05_quad_with_dynamics	0.48	Ouster OS1-64
ncd_4	06_dynamic_spinning	0.09	Ouster OS1-64
ncd_5	07_parkland_mound	0.70	Ouster OS1-64
ulhk_1	HK-Data20190117	0.60	Velodyne HDL-32E
ulhk_2	HK-Data20190426-2	0.74	Velodyne HDL-32E
bg_1	1006-01	0.76	Velodyne VLP-16 & Livox AVIA
bg_2	1008-03	0.74	Velodyne VLP-16 & Livox AVIA
avia_1	hku_main_buiding	0.96	Livox AVIA
avia_2	outdoor_Mainbuilding_100Hz_2020-12-24-16-46-29	0.14	Livox AVIA
avia_3	outdoor_run_100Hz_2020-12-27-17-12-19	0.09	Livox AVIA

4. Mapping Results

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

NCLT 2012-05-11

Newer College Dataset 02_long_experiment

hku_main_building

5. Paper

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}}

6. Acknowledgements

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

7. Known Issues

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.

8. Time Line

Time	Event
Aug 13, 2023	😀 Paper submitted to IEEE Robotics and Automation Letters (RA-L)
Nov 5, 2023	😭 Revise and resubmit
Dec 22, 2023	🥳 Paper accepted for publication in RA-L
Current	🎉 Source code released