Automatic multi-LiDAR calibration and refinement method

Pengjin Wei, Guohang Yan, Yikang Li, Kun Fang, Xinyu Cai, Wei Liu, Jie YangMeasurement 2024 · DOI

TL;DR

This paper introduces CROON, a novel two-stage method for automatically calibrating multiple LiDAR sensors in road scenes. CROON achieves high accuracy and robustness by first performing a rough calibration using road scene characteristics and then refining the poses with an octree-based optimization, making it a significant step towards reliable autonomous driving perception.

What Problem Does This Paper Address?

Autonomous driving systems rely heavily on accurate environmental perception, which is often achieved by equipping vehicles with multiple LiDAR sensors to cover a wide field of view and capture dense point clouds. However, the effectiveness of these systems hinges on precise calibration of the extrinsic parameters (rotation and translation) between these LiDARs. Traditional calibration methods are often manual, require specific targets, or are sensitive to motion drift, making them inefficient, error-prone, and difficult to apply in large-scale, real-world scenarios. This paper addresses the need for an automatic, robust, and accurate method for multi-LiDAR calibration that can handle arbitrary initial poses and is independent of external targets.

Methods

CROON employs a two-stage approach:

  1. Rough Calibration: This stage leverages the inherent characteristics of road scenes, specifically the prevalence of ground planes. The method first identifies and aligns ground planes from different LiDARs to estimate pitch, roll, and vertical translation. It then uses a simplified cost function to efficiently determine yaw and horizontal translations. This stage is designed to quickly constrain the LiDAR poses from large initial deviations.
  2. Refinement Calibration: Following the rough calibration, this stage iteratively refines the LiDAR poses. It utilizes an Iterative Closest Point with Normal (ICPN) variant, which enhances point features by incorporating normal information to improve robustness. This is further optimized using an octree-based method. This octree approach recursively subdivides space, measuring the volume occupied by point clouds within these subdivisions to minimize the overall space volume, thereby achieving accurate alignment.

Datasets

The paper utilizes two types of datasets:

  • Real-world data: Collected from a driverless vehicle platform equipped with three LiDARs. This dataset is used for qualitative evaluation and consistency analysis.
  • Simulated data: Generated using the CARLA simulator, which provides ground truth for quantitative analysis. This dataset includes more LiDARs and is used to demonstrate robustness, stability, and accuracy.

The authors also mention that all related datasets and codes are open-sourced.

Key Insights and Ideas

  • Leveraging Road Scene Geometry: The core innovation lies in exploiting the ubiquitous ground plane in road scenes for initial, robust calibration. This makes the method "target-less" and applicable in diverse environments.
  • Two-Stage Calibration for Efficiency and Accuracy: The combination of a fast rough calibration followed by a precise refinement stage is crucial. The rough stage handles large initial errors, while the refinement stage achieves high accuracy without being susceptible to local optima.
  • Octree-Based Refinement: The octree method provides an intuitive and effective way to measure and minimize the spatial volume occupied by aligned point clouds, leading to precise pose estimation.
  • ICPN for Enhanced Point Features: Incorporating point normals into the ICP process significantly improves its ability to handle sparse point clouds and extract richer geometric information.

Key Results

  • High Accuracy: CROON achieves good results in all six degrees of freedom (pitch, roll, yaw, x, y, z), with mean errors often in the order of hundredths of a degree or millimeter.
  • Robustness and Stability: The method demonstrates consistency and stability across multiple scenes and initial configurations, as evidenced by low standard deviations in error measurements.
  • Efficiency: The two-stage approach, particularly the rough calibration, is designed to be efficient, handling large initial deviations quickly.
  • Comparison: CROON outperforms other targetless methods in terms of accuracy for several degrees of freedom, as shown in Table II.
  • Qualitative Improvements: Visualizations (Figure 2) clearly show the progressive alignment of LiDAR point clouds from an initial scattered state to a well-fused representation.

Limitations and Future Work

  • Failure Cases: The paper acknowledges that the method may fail in extremely empty or simple road scenes, or when scenes are very narrow, as the point clouds from LiDARs might become "useless signals."
  • Future Work: While not explicitly detailed, the authors' focus on robustness and accuracy in diverse road scenes suggests potential for extending this to more complex environments or integrating it into a full autonomous driving perception pipeline. The open-sourcing of code and data also facilitates further research and development by the community.

Summary updated

Abstract

Abstract Sensor-based environmental perception is a crucial component of autonomous driving systems. To perceive the surrounding environment better, an intelligent system would utilize multiple LiDARs (3D Light Detection and Ranging). The accuracy of the perception largely depends on the quality of the sensor calibration. This research aims to develop a robust, fast, automatic, and accurate calibration strategy for multiple LiDAR systems. Our proposed multi-LiDAR calibration method consists of two stages: rough and refinement calibration. In the first stage, sensors are roughly calibrated from an arbitrary initial position using a deep neural network that does not rely on prior information or constraints on the initial sensor pose. In the second stage, we propose the octree-based refinement, an optimization method that considers sensor noise and prioritization. Our strategy is robust, fast, and not restricted to any environment. Additionally, we collected two datasets consisting of both real-world and simulated scenarios. Our experimental results from both datasets demonstrate the reliability and accuracy of our method. All the related datasets and codes are open-sourced on the GitHub website at https://github.com/OpenCalib/LiDAR2LiDAR .
BibTeX
@inproceedings{wei2024automatic,
  author = {Pengjin Wei and Guohang Yan and Yikang Li and Kun Fang and Xinyu Cai and Wei Liu and Jie Yang},
  title = {{Automatic multi-LiDAR calibration and refinement method}},
  booktitle = {Measurement},
  year = {2024},
  doi = {10.1088/1361-6501/ad3c60},
}
Figure