$\color{#FF577E}{𝙁𝙧𝙚𝙚\text{-}𝙄𝙣𝙞𝙩}$
Scan-Free, Motion-Free, and Correspondence-Free Initialization for Doppler LiDAR-Inertial Systems

1. Introduction

$\color{#FF577E}{𝙁𝙧𝙚𝙚\text{-}𝙄𝙣𝙞𝙩}$ is an initialization framework for 4D Doppler LiDAR-inertial systems that eliminates the need for LiDAR scan motion compensation (undistortion), excitation motions, and feature or map correspondences by fusing point-wise Doppler velocity measurements with inertial data under non-inertial kinematics. $\color{#FF577E}{𝙁𝙧𝙚𝙚\text{-}𝙄𝙣𝙞𝙩}$ is also plug-and-play compatible with typical LiDAR-inertial systems, and can handle initialization both in arbitrary environments, including structure-degenerated scenes, and under a wide range of initial motions, from stationary to dynamic and even aggressive. The proposed Doppler-Inertial Velocimeter (DIV) enables fast convergence and high-rate estimation, delivering estimator outputs over 10 kHz.

$\color{#FF577E}{𝙁𝙧𝙚𝙚\text{-}𝙄𝙣𝙞𝙩}$ provides accurate initial-state estimation in both structured and structure-degenerated environments, as its intrinsic correspondence-free nature do not depend on environmental structures. The following results demonstrate the initialization performance of $\color{#FF577E}{𝙁𝙧𝙚𝙚\text{-}𝙄𝙣𝙞𝙩}$ in structured and structure-degenerated scenes.

Framework Highlights

• 🚀 Scan-Free: Point-wise scheme. No requirement of motion compensation (undistortion) of LiDAR scans.
• 🧭 Motion-Free: No requirement of excitation motions or specific motion assumptions.
• 🔗 Correspondence-Free: No requirement of feature or map correspondence.
• ⚡ High-Rate: The Doppler-Inertial Velocimeter (DIV) outputs estimation results over 10 kHz.
• 🔌 Plug-and-Play: Easily integrated with typical LiDAR-inertial systems.

Beyond the initialization of Doppler LiDAR-inertial systems, $\color{#FF577E}{𝙁𝙧𝙚𝙚\text{-}𝙄𝙣𝙞𝙩}$ pinpoints the crux of general initialization problems: once velocity or differential signal is directly observable, initialization no longer relies on motion compensation, motion excitation, motion assumption, or correspondences. This logic and structural design are sensor-agnostic and carry over to other systems, such as visual-inertial odometry (VIO), whenever a sensor or a learned model provides differential, velocity-aware measurements or predictions.

2. Source Code of $\color{#FF577E}{𝙁𝙧𝙚𝙚\text{-}𝙄𝙣𝙞𝙩}$

$\color{#FF577E}{𝙁𝙧𝙚𝙚\text{-}𝙄𝙣𝙞𝙩}$ is seamlessly integrated into the $\color{#F1261D}{𝗙}\color{#FF9200}{𝗠}\color{#FDDB2A}{𝗖}\color{#4DE42B}{𝗪}\color{#0FF692}{\text{-}}\color{#0BD8F9}{𝗟}\color{#2C22FB}{𝗜}\color{#C32DF3}{𝗢}$ (DOI | Project Page) framework. Please refer to the initialization module of $\color{#F1261D}{𝗙}\color{#FF9200}{𝗠}\color{#FDDB2A}{𝗖}\color{#4DE42B}{𝗪}\color{#0FF692}{\text{-}}\color{#0BD8F9}{𝗟}\color{#2C22FB}{𝗜}\color{#C32DF3}{𝗢}$ (DOI | Project Page) for the source code of $\color{#FF577E}{𝙁𝙧𝙚𝙚\text{-}𝙄𝙣𝙞𝙩}$.

3. Free-Init Dataset

The dataset for qualitative test of dynamic initialization (Free-Init Dataset) is available on Hugging Face and Google Drive. The characteristics of each sequence are summarized in the following table.

Sequence	Preview	FMCW-LIO Map	Platform	Distance (m)	Duration (s)
Free-Init_handheld_campus			Handheld	30	30
Free-Init_handheld_tunnel			Handheld	50	30
Free-Init_vehicular_highway			Vehicular	553	30
Free-Init_vehicular_tunnel			Vehicular	562	30

4. Citation

If you find our algorithms, datasets, or frameworks helpful in your research, please consider citing our works.

@article{zhao2024free-init,
  title={Free-Init: Scan-Free, Motion-Free, and Correspondence-Free Initialization for Doppler LiDAR-Inertial Systems},
  author={Zhao, Mingle and Wang, Jiahao and Gao, Tianxiao and Xu, Chengzhong and Kong, Hui},
  journal={IEEE Robotics and Automation Letters},
  volume={9},
  number={12},
  pages={11329--11336},
  year={2024},
  publisher={IEEE}
}

@article{zhao2024fmcw-lio,
  title={FMCW-LIO: A Doppler LiDAR-Inertial Odometry},
  author={Zhao, Mingle and Wang, Jiahao and Gao, Tianxiao and Xu, Chengzhong and Kong, Hui},
  journal={IEEE Robotics and Automation Letters},
  volume={9},
  number={6},
  pages={5727--5734},
  year={2024},
  publisher={IEEE}
}