Objective Insufficient illumination, significant dust interference, and sparse point clouds in underground coal mine working faces impede the stable extraction of features, thereby severely limiting the accuracy and robustness of camera-LiDAR joint calibration based on unstructured scene characteristics.

Methods This paper addresses complex operational scenarios, including fully mechanized mining faces and excavation roadways, by proposing a joint calibration method that leverages cross-depth feature coupling between LiDAR and camera. For feature extraction, an enhanced RANSAC multi-plane fitting algorithm is introduced, integrating normal vector pre-clustering and an adaptive iteration mechanism to efficiently extract geometric structures, such as hydraulic support roof beams and roadheader casings. A cross-depth edge fusion strategy is also proposed, which synergistically utilizes curvature discontinuities and planar intersection features to improve the completeness and robustness of edge structures. The calibration framework employs a two-stage registration strategy: coarse registration rapidly estimates initial extrinsic parameters via an axial cyclic perturbation approach, while fine registration establishes joint point-line constraints within Lie group space and iteratively refines them through nonlinear optimization. This ensures high-precision alignment is achieved even under challenging conditions involving coal dust and complex operational environments.

Results and Conclusions  The proposed method was validated both on the Gazebo simulation platform and in real underground mining scenarios. Experimental results show that under noise-free conditions, the rotational error is less than 0.2°, the translational error is below 0.02 m, and the mean reprojection error does not exceed 3.5 px. Even in high-noise environments, the method still maintains excellent stability. Compared with traditional approaches, the proposed method achieves mean reprojection errors of 2.89 px and 3.03 px on the roadheader face and longwall face, respectively, demonstrating a significant improvement. Moreover, the method does not rely on manually placed calibration targets and exhibits strong environmental adaptability and robustness, meeting the high-precision calibration requirements of multi-modal perception systems in complex underground coal-mine environments.