2024 Volume 80 Issue 22 Article ID: 23-22014
Road administrators conduct road surface condition surveys and visual inspections to monitor the condition of road pavements. However, they face challenges in terms of labor and cost. Therefore, technology that uses deep learning to analyze video images has been attracting attention. These images of road pavements are captured by drive recorders and smartphones. In existing research, the accuracy of the training data is low, and shadows on the road surface reduce the detection accuracy of cracks.
In this study, we developed a deep learning model designed based on specific characteristics and methodologies derived from the results of a road surface condition survey. Additionally, we devised a simple method to calculate crack rates by removing shadows from 4K-resolution video images using machine learning. Through demonstration experiments, we found that the crack ratio from the road surface condition survey results had an accuracy of approximately 60%. This is equivalent to the accuracy of visual inspections, suggesting that cracks in road pavements can be easily inspected.
