Deep Learning and Reality Capture for Automated Railway Bridge Inspection with BIM Enrichment

Abstract: Manual inspection remains subjective and poorly integrated with digital workflows. Although computer vision enables automated damage detection, most approaches fail to translate observations into parametric Building Information Modelling (BIM). For bridges, this limitation persists within Bridge Information Modelling (BrIM), where geometry, semantics, and lifecycle information must align. This paper addresses the gap through a four-stage framework: optimal reality modelling, hybrid damage perception, automated Scan-to-BIM, and semantic enrichment via IFC objects. Validated on operational railway bridges, the method achieves a damage-detection mAP50 of 0.621 and geometric deviations under 23 mm, enabling BrIM generation directly from visual data.