Core Concepts
A robust and accurate online, target-free LiDAR-camera extrinsic calibration approach that leverages state-of-the-art large vision models for cross-modal mask matching.
Abstract
The article presents MIAS-LCEC, an online, target-free LiDAR-camera extrinsic calibration (LCEC) approach that employs a novel coarse-to-fine strategy to accurately estimate the extrinsic parameters. The key components are:
A virtual camera is introduced to project the LiDAR point cloud into a LiDAR intensity projection (LIP) image, which is then aligned with the RGB image.
Both the LIP and RGB images are segmented using MobileSAM, a state-of-the-art large vision model, to extract informative features.
A novel cross-modal mask matching (C3M) algorithm is developed to generate sparse yet reliable correspondences, which are then propagated to obtain dense matches.
The dense correspondences are used as inputs for a perspective-n-points solver to derive the extrinsic matrix.
The authors also provide a versatile LCEC toolbox with an interactive visualization interface, and publish three real-world datasets to comprehensively evaluate the performance of LCEC algorithms. Extensive experiments demonstrate that MIAS-LCEC outperforms state-of-the-art online, target-free approaches, particularly in challenging scenarios, and achieves similar performance to an offline, target-based algorithm.
Stats
The mean rotation error is reduced by 22-88% and the mean translation error is decreased by 40-95% compared to existing state-of-the-art algorithms.
Quotes
"Our main contributions are threefold: we introduce a novel framework known as MIAS-LCEC, provide an open-source versatile calibration toolbox with an interactive visualization interface, and publish three real-world datasets captured from various indoor and outdoor environments."
"The cornerstone of our framework and toolbox is the cross-modal mask matching (C3M) algorithm, developed based on a state-of-the-art (SoTA) LVM and capable of generating sufficient and reliable matches."