CulvertVision: Advancing AI-Augmented Inspection and Condition Assessment of Culverts

Culvert inspection remains a persistent bottleneck in infrastructure asset management, constrained by labor-intensive manual inspection, resource limitations, and varying reporting across varied environments. This study presents CulvertVision, a deep learning framework for automated defect detection in culvert inspection videos designed to align with the Pipeline Assessment and Certification Program (PACP) standards. Utilizing a ResNet-101 architecture fine-tuned via transfer learning from sewer infrastructure benchmarks, the system processes untrimmed inspection videos to generate temporally stable classifications for four categories: corrosion, settled deposits, joint offset, and normal. Evaluation across 6,600 annotated frames under shared-video, noise-augmented, and video-disjoint configurations reveals that while noise augmentation improves robustness in controlled settings ( F 1 - scores > 0.88 ), a significant generalization gap emerges in out-of-distribution scenarios, with joint offset performance declining to an F 1 -score of 0.55. Beyond accuracy, CulvertVision’s throughput analysis demonstrates high operational efficiency, offering substantial time savings compared to manual human review. By automating PACP-aligned defect detection, the framework enables agencies to streamline inspection workflows and support data-driven capital and maintenance planning. While current work explores dataset expansion through generative augmentation to address training limitations, CulvertVision represents a scalable prototype for artificial intelligence (AI)–augmented infrastructure inspection and broader field deployment.