Bridge Inspection Support Framework Incorporating Human Reliability

Abstract

In Japan, periodic road-bridge inspections rely primarily on inspectors’ close visual examination at five-year intervals. Concurrently, the proportion of aging bridges is increasing while the availability of experienced inspectors is expected to decline, amplifying the demand for support tools that reduce workload without compromising safety assurance. This study proposes a bridge-deck inspection support framework grounded in human–AI collaboration and presents its core application for crack-candidate visualization on concrete decks. The framework enforces an explicit division of roles: the AI is restricted to presenting and visualizing crack candidates, whereas inspectors retain responsibility for close visual inspection, on-site measurement, and final condition assessment. As the AI component, we developed a semantic-segmentation model that estimates crack presence at the pixel level and implemented post-processing to suppress short, spurious detections while preserving line-like crack structures. Evaluation results indicate that missed detections of cracks as continuous line-like structures were rare in our test set, including images with diverse surface appearances. False positives occurred predominantly as short segments, which—while requiring verification—are typically less hazardous than misses in safety-critical inspections and can be treated as conservative prompts for closer examination. Accordingly, this study frames crack detection not solely as an accuracy problem but as a human reliability intervention: a “safety-oriented filter” that redistributes attention from exhaustive search to prioritized verification, thereby mitigating lapse-type inspection errors. The findings suggest that the proposed framework can support safer attention allocation while maintaining inspector accountability.

Keywords: Bridge Inspection, Human–AI Collaboration, Semantic Segmentation, Cognitive Load, Attention Guidance, Fail-safe Mechanism, Human Reliability

DOI: 10.54941/ahfe1007966