2026 Volume 7 Issue 3 Pages 199-207
Regular assessment of crack development is essential for maintaining railway concrete structures. Yet, current inspections rely heavily on close visual observation, which is prone to subjective judgment, oversight, and considerable recording workload. As the shortage of experienced inspectors worsens, automated crack detection using image analysis is increasingly required. While several studies have been conducted from this perspective, most assume frontal shooting with sufficient illumination and camera‑to‑surface distance, making them difficult to apply in narrow, close‑range, and low‑light environments. In this study, we conducted a foundational trial on a mock RC structure on a university campus, integrating video recording, photogrammetric orthophoto generation, and AI-based crack detection under constrained, low-light conditions. In parallel, we are conducting data acquisitions in an actual railway tunnel. Through investigations under various conditions, we set shooting conditions that enable the generation of high-precision 3D models and orthophotos without geometric distortion. We quantitatively evaluated the impact of recording speed on crack detection accuracy. Despite a tendency towards overdetection, recording at 1.0 km/h yielded higher recall and F-value than recording at 2.0 km/h. This was addressed by prioritizing recall, assuming that false detections could be eliminated in a subsequent manual verification workflow. These results suggest that the proposed video-photogrammetry workflow could form the basis of a practical diagnostic method for railway concrete structures.