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

The rapid advancement of autonomous driving and embodied intelligence highlights the critical need for robust multi-sensor systems. In such systems, LiDAR and cameras are frequently paired-leveraging their complementary strengths: LiDAR excels at high-precision measurements, while cameras deliver rich texture information. Accurate extrinsic calibration between these two sensors is a fundamental prerequisite for effective data fusion. However, traditional calibration methods face challenges in extracting robust multi-modal features under complex scenarios and typically overlook feature visibility check during optimization. To address these limitations, we propose a novel target-free method for the automatic extrinsic calibration of LiDAR and camera. Our approach incorporates key innovations in both feature extraction and pose optimization. For LiDAR point clouds, we design a feature detector based on local eigenvalue decomposition to reliably identify geometric edge points. For images, we overcome the limitations of traditional line feature extraction methods, which are susceptible to texture interference, by fusing raw images with depth maps estimated via Depth Anything V2, enabling robust extraction of geometric edge lines. Furthermore, to mitigate the impact of invisible features, a step often ignored in existing methods, we integrate a visibility check module into our optimization framework. Finally, LiDAR-camera extrinsic calibration is achieved by establishing correspondences between the extracted multi-modal geometric edge features and minimizing the associated constraints. Comprehensive experiments across diverse scenes show that our method consistently outperforms state-of-the-art (SOTA) counterparts. Notably, even in the worst-case scene, it achieves rotational errors below 0.86° and translational errors below 0.09 meters, outperforming the top existing methods by a substantial margin. For full reproducibility, our code is publicly accessible at https://cslinzhang.github.io/CalibrationLC.