Abstract
This study addresses the challenge of intelligent detection and spatial localization of massive cracks in deep-cut canal sections of the Middle Route of China’s South-to-North Water Diversion Project. An integrated framework for crack detection, localization, and three-dimensional visualization of concrete canal linings is proposed. UAV-based imitation ground photogrammetry was first employed to acquire 114,220 high-resolution images, from which a photogrammetric textured mesh model with a surface resolution of 0.47 cm/pixel was constructed. Based on 2886 representative images, a dedicated training dataset for intelligent crack detection was established. By integrating photogrammetric collinearity equations with real-time position and orientation system (POS) data from Unmanned aerial vehicle (UAV) imitation ground flights, a single-image crack coordinate calculation model was developed and embedded into the YOLOv7 object detection framework. This integration enables direct computation of crack spatial coordinates from a single image without reliance on stereo image pairs, allowing crack identification and spatial localization to be synchronously achieved within a unified deep learning framework. Experimental results show that YOLOv7 achieves an mAP@0.5 of 84.3% at an IoU threshold of 0.5, and the planar localization accuracy is better than 0.1 m. Finally, the detected and localized cracks are mapped onto the millimeter-level photogrammetric textured mesh model, enabling intuitive visualization of crack spatial distribution and providing technical support for structural condition assessment and intelligent operation and maintenance of long-distance water conveyance channels.