Towards High-Precision Target-Free LiDAR-Camera Extrinsic Calibration: Multi-Modal Geometric Edge Matching and Visibility-Guided Optimization

Baosheng Zhang Lin Zhang^* Shengjie Zhao Yicong Zhou

Tongji University

We propose a novel target-free method for the automatic extrinsic calibration of LiDAR and camera to address the limitations of traditional approaches. Our key innovations lie in robust multi-modal feature extraction and an optimized process. For point clouds, we employ feature detectors based on local eigenvalue decomposition to identify geometric edge points. For images, we overcome texture interference by fusing raw images with learning-based depth estimation to extract geometric edge lines. Furthermore, we incorporate a critical visibility check module into the optimization to mitigate the influence of invisible features—a step often neglected in existing methods. The extrinsic parameters are then computed by establishing correspondences between these geometric edge features and minimizing a constraint model. Its key innovations include:

A Novel Point Cloud Feature Extraction Algorithm: A novel point cloud feature extractor is proposed, which employs the decomposition of local points to effectively identify and acquire salient edge points from the point cloud.

A Novel Image Feature Extraction Algorithm: This algorithm achieves the extraction of 3D edge features from 2D image by integrating traditional 2D image feature detection with a learning-based depth estimation strategy.

A Novel Visibility Check Module: This module assesses the visibility of each point by analyzing the Gaussian ellipsoid overlapping and the normal vector consistency, effectively identifying and removing the influence of occluded points during optimization.

Framework

After independent extraction of geometric edge features from both point clouds and image, a visibility analysis module filters out occluded points; an elliptical matching strategy then establishes cross-modal correspondences, followed by optimization of extrinsic parameters [R, t] through reprojection error minimization.

Experiments

To validate the performance of our proposed method, we collected data from multiple real-world scenes using our experimental setup. Each scene was recorded statically for approximately 10 seconds to ensure the acquisition of dense point clouds. All experiments were conducted on a laptop with an Intel Core i7-10510U CPU@1.80GHz x 8 and 32 GB of RAM, without GPU acceleration. For comparative analysis, we evaluated our method against several SOTA calibration methods: Pandey, Yuan , Koide, and MFCalib.

Center Image in Figure

Our LiDAR-camera experimental setup, featuring a Livox-AVIA LiDAR and a HIKROBOT MV-CA050-12UC camera.

The Performance of the Visibility Check Module

We have conducted a qualitative experimental analysis of our proposed feature visibility analysis model on diverse point cloud datasets, and present the results below.

Comparative Experiment with the SOTA Methods

We conducted comparative experiments between our method and several state-of-the-art calibration approaches (Pandey, Yuan , Koide, and MFCalib) across multiple challenging indoor and outdoor scenes. The results presented below demonstrate that our method achieves outstanding robustness and accuracy in these demanding environments.