Radar4Motion: IMU-Free 4D Radar Odometry with Robust Dynamic Filtering and RCS-Weighted Matching

Radar4Motion is a robust odometry method that utilizes Doppler and RCS information from the 4D imaging radar's point cloud, even in the presence of noisy and sparse point cloud data.

View-of-Delft Dataset Seq 03

View-of-Delft Dataset Seq 17

• The above gif shows ONLY odometry-based mapping results.
  • NO inertial sensor, NO GNSS sensor, NO loop-closure
  • Only Single front-view 4D Imaging Radar!

Prerequisites

To run this project, you need:

• ROS (Robot Operating System), tested with ROS Noetic
• Eigen3, for matrix and vector operations
• PCL (Point Cloud Library), for handling radar point cloud data
• nlohmann_json, for JSON parsing

Installation

Clone the Repository

cd ~/catkin_ws/src
git clone https://github.com/ailab-hanyang/Radar4Motion.git

Build the Project

Using catkin, set up your workspace and build the project:

cd ~/catkin_ws
catkin_make
source devel/setup.bash

Using the View-of-Delft (VoD) Dataset

If you want to evaluate Radar4Motion with the View-of-Delft (VoD) dataset, please do the following:

1. Download or place the VoD dataset in your preferred location.

2. Set VoD datset path str_vod_dataset_base_path in /launch/radar4motion_offline.launch.

   • OR, create a symbolic link inside the dataset/ folder so the code can locate VoD data under dataset/view_of_delft_PUBLIC.

   • Set <param name="str_vod_dataset_base_path" value="$(find radar4motion)/dataset/"/>

   • For example:

     cd ~/catkin_ws/src/Radar4Motion
     mkdir -p dataset
     ln -s /path/to/view_of_delft_PUBLIC dataset/

     Folder tree:

     Radar4Motion/
     ├── src/
     │   ├── .cpp files..
     └── dataset/
         └── clips/
         └── view_of_delft_PUBLIC/  -> symlink to /path/to/view_of_delft_PUBLIC
             ├── radar/
             │   ├── training/
             │   │   ├── velodyne/
             │   │   │   ├── (bin files)
             │   │   ├── pose/
             │   │   │   ├── (label files)

3. Create a test folder for evaluation (option)

   cd ~/catkin_ws/src/Radar4Motion
   mkdir -p test

Usage

To start the radar odometry processing, launch the provided ROS launch files:

• View-of-delft dataset

  roslaunch radar4motion radar4motion_offline.launch

• ROS topic

  roslaunch radar4motion radar4motion_online.launch

• Note: The accumulated scans-to-submap matching algorithm requires precise sensor-vehicle calibration.
  Therefore, you must accurately set the following parameters in the ./config/radar_point_cloud_odometry.ini file according to your environment:

  m_d_radar_calib_x_m = 3.5
  m_d_radar_calib_y_m = 0.0
  m_d_radar_calib_z_m = 0.0
  m_d_radar_calib_roll_deg = 1.0
  m_d_radar_calib_pitch_deg = -0.568
  m_d_radar_calib_yaw_deg = 0.43

  • In particular, for the View-of-Delft (VOD) dataset, exact sensor-vehicle calibration values are not provided by the dataset.
  • Hence, we have set approximate values based on the sensor mounting position and vehicle images. These values will be updated in the future through our ongoing radar4selfcalibration research, which will estimate these parameters more accurately.

Configurations

Detailed descriptions of parameters can be found in PARAMETERS.

1. ROS & File path configuration
   • File: /launch/radar4motion_[offline/online].launch
2. Radar ego motion estimation (ref: REVE)
   • File: /config/radar_ego_motion_estimation.ini
3. Odometry
   • File: /config/radar_point_cloud_odometry.ini

TODO

• Update voxel struct (frame-scan unit)
• Update VoD calibration params with Radar4SelfCalibration

License

This project is licensed under the Apache-2.0 License - see the LICENSE.md file for details.

Authors

• Soyeong Kim - [email protected]
• Jiwon Seok - [email protected]

Citation

If you find this work useful, please cite the following paper:

@article{kim2024radar4motion,
title={Radar4Motion: IMU-Free 4D Radar Odometry with Robust Dynamic Filtering and RCS-Weighted Matching},
author={Kim, Soyeong and Seok, Jiwon and Lee, Jaehwan and Jo, Kichun},
journal={IEEE Transactions on Intelligent Vehicles},
year={2024},
publisher={IEEE}
}

Acknowledgement

We would like to express our gratitude to all the contributors and resources that made this research possible.

• In the development of this package, we refer to KISS-ICP and REVE for source codes.
• Dataset: View-of-Delft (VoD)
• Evaluation: evo package for odometry evaluation