Automated Interpretation of Culvert Inspection Videos Using AI and Computer Vision

Details

• Creators:
  Mohammadi, Pouria ; Rashidi, Abbas
• Corporate Creators:
  University of Utah. Department of Civil and Environmental Engineering
• Corporate Contributors:
  Utah Department of Transportation ; United States. Department of Transportation. Federal Highway Administration
• Subject/TRT Terms:
  Artificial Intelligence Computer Vision Culverts Information Processing Inspection Video
• Publication/ Report Number:
  UT-25.20
• Resource Type:
  Tech Report
• Geographical Coverage:
  United States ; Utah
• Edition:
  Final August 2023 to August 2025
• Corporate Publisher:
  Utah Department of Transportation
• Abstract:
  Culvert assets play a critical role in ensuring the safe operation of highways. UDOT maintains over 120,000 drainage culverts and storm drain pipes along state highways. To maintain these assets optimally and prevent failures, UDOT must collect comprehensive information about all culverts across the state and inventory them in the ATOM system. Accurate identification of culverts needing repair, rehabilitation, or replacement necessitates thorough and well-documented inspections. Traditional culvert inspection is slow and prone to subjectivity, leading to inconsistencies in the assessment of culvert conditions. This project focused on automating the interpretation of culvert inspection videos using advanced computer vision and deep learning techniques. Beginning with a small and imbalanced dataset, the team expanded the data through additional data collection and augmentation, followed by manual labeling and annotation of structural defects. Three model types were developed to support different stages of inspection analysis: a binary classification model to identify defective frames, multiclass image classification models to classify five major defect types, and an object detection model capable of localizing and classifying defects. To bridge the gap between model output and practical deployment, graphical user interfaces (GUIs) were created for each model type, enabling UDOT staff to analyze inspection videos, receive condition ratings, and generate detailed summary reports without technical expertise. When tested on 56 real-world videos, the object detection GUI correctly assessed culvert conditions in 84% of cases. The system offers a scalable, cost-effective, and objective approach to culvert inspection, reducing manual workload and increasing the consistency and accuracy of infrastructure condition assessments.
• Format:
  PDF
• Funding:
  24-8333
• Collection(s):
  US Transportation Collection
• Main Document Checksum:
  urn:sha-512:75b358fcf63081cd94a3e98c65a8a0ee0477751ab9639ce2fa7e8d10ed95f6823db893896dba75ab6cf659e927e78c28c2fa3577e985d88b18796109dfa1c677
• Download URL:
  https://rosap.ntl.bts.gov/view/dot/88332/dot_88332_DS1.pdf
• File Type:
  [PDF - 3.82 MB ]