3D Rock Fracture Characterization Based on Monocular Depth Estimation and Instance Segmentation

Addressing inefficiency, safety hazards, and environmental adaptability limitations in existing rock fracture measurement methods, this paper proposes a 3D characterization method for rock fractures integrating monocular depth estimation with instance segmentation. The approach achieves non-contact measurement and automatic extraction of 3D geometric parameters (length, width, orientation) using standard monocular cameras. An enhanced Dense Prediction Transformer (DPT) generates high-resolution depth maps and reconstructs 3D point clouds from single images. Mask R-CNN performs pixel-accurate fracture segmentation to isolate fracture point clouds. Based on extracted point clouds, this work develops a 3D parameter quantification algorithm: (1) Fracture length calculation via medial-axis-transform skeleton extraction and Euclidean distance accumulation; (2) Width measurement combining local principal-direction projection with bidirectional nearest-neighbor search; (3) Orientation determination through least-squares fitting of skeleton traces. Experimental results validate the method's efficacy: Average relative errors for length/width are 8.73%/10.29%; mean absolute errors for dip/direction are 3.92°/7.92°, meeting engineering accuracy requirements. The method provides reliable technical support for rapid, safe, and automated 3D digital characterization in complex engineering environments.