Ground-vision toolkit

A toolkit for ground feature processing, inverse perspective mapping (IPM) and so on, which is applied in our preprint Ground-VIO.

News

• [2023/07/11] - Initial release of code and dataset.

Introduction

This toolkit fully utilizes the camera-ground geometry for accurate ground feature tracking. In the preprint, we verify that the estimation of vehicle states, the calibration of camera-ground geometry and stable feature tracking could be leveraged in a monocular visual-inertial estimator.

Dataset

This repo also provides the urban road visual-inertial dataset used in Ground-VIO.

The dataset contains both Carla simulated data sequences (S-A, S-B) and real-world data squences (R-A, R-B, R-C, R-D, R-E and R-F). The detail information is listed below. For real-world data sequences, the GT poses are obtained from the forward-and-backward smoothed solution of PPK/tactical-grade IMU integration.

Simulation Data (Carla)

Sequence	Date	Length	Sensors	Features	Image
S-A	-	110 s	IMU/Camera/Semantic	Urban road
S-B	-	135 s	IMU/Camera/Semantic	Urban road

Real-World Data

Sequence	Date	Length	Sensors	Features	Image
R-A	2022/10/12	180 s	IMU/Camera	Urban road
R-B	2022/10/12	180 s	IMU/Camera	Urban road
R-C	2022/10/12	180 s	IMU/Camera	Urban road
R-D	2022/10/12	180 s	IMU/Camera	Urban road
R-E	2022/10/12	270 s	IMU/Camera	Highway
R-F	2022/10/12	270 s	IMU/Camera	Highway

Dataset is available at OneDrive.

Dependencies

The dependencies include Eigen and OpenCV. We use the camodocal project to handle camera models, while we modify it to a minimal version which doesn't need Ceres.

Getting Started

The project could be built either with or without ROS.

Building with ROS

Follow the steps to build the project in a ROS workspace

mkdir catkin_ws
mkdir catkin_ws/src
cd catkin_ws/src
git clone https://github.com/GREAT-WHU/gv_tools
cd ..
catkin_make

To run the ground tracker node, following

source devel/setup.bash
roslaunch gv_tools track_carla_example.launch

and a rviz viewer would be simultaneously launched.

Then, play the data bag in another terminal

rosbag play s_a.bag

Building without ROS

To build the project just as a plain CMake project

git clone https://github.com/GREAT-WHU/gv_tools
cd gv_tools
mkdir build && cd build
cmake ..
make -j8

Run the example track_dataset following

./build/track_dataset ./config/realworld/tracker.yaml DATASET_DIR/data_r_a/cam0/ DATASET_DIR/data_r_a/stamp.txt DATASET_DIR/data_r_a/gt_pose.txt 

Acknowledgements

The toolkit is developed by GREAT (GNSS+ REsearch, Application and Teaching) Group, School of Geodesy and Geomatics, Wuhan University.

Thanks to VINS-Fusion for excellent open-source codes. Thanks to msckf-vio for inspring grid-based feature extraction method.

Credit / Licensing

@article{zhou2023groundvio,
  title={Ground-VIO: Monocular Visual-Inertial Odometry with Online Calibration of Camera-Ground Geometric Parameters},
  author={Yuxuan Zhou, Xingxing Li, Shengyu Li, Xuanbin Wang, Zhiheng Shen},
  journal={arXiv preprint arXiv:arXiv:2306.08341},
  year={2023}
}

The codebase and documentation is licensed under the GNU General Public License v3 (GPL-3).