A physic-guided YOLO framework for pavement deformation distress detection.

Area of Science:

• Civil Engineering
• Computer Science
• Artificial Intelligence

Background:

• Pavement distress, such as rutting and corrugation, significantly impacts road safety and maintenance costs.
• Automated detection methods are needed to efficiently assess pavement conditions.

Purpose of the Study:

• To develop an automated methodology for detecting and assessing pavement deformation distress using deep learning.
• To enhance the accuracy and reliability of pavement distress detection models.

Main Methods:

• Exploration of various YOLO algorithms for pavement distress detection, with YOLOv5 showing initial promise.
• Development of a dedicated YOLOv5 model with auto-annotation for corrugation detection due to limited data.
• Integration of a Physics-Informed Neural Network (PINN) into the YOLO framework to create a hybrid PINN-YOLO model.

Main Results:

• The baseline YOLOv5 model achieved high mean Average Precision (mAP) for rutting (96.3% high severity, 81.9% low severity) and 66.5% mAP for corrugation.
• The hybrid PINN-YOLO model demonstrated statistically significant improvements in bounding box accuracy and reduction in false positives for both rutting and corrugation.
• The hybrid model yielded superior mAP results compared to the baseline YOLOv5 models, indicating enhanced prediction reliability.

Conclusions:

• The hybrid PINN-YOLO model offers a robust and accurate solution for automated pavement distress evaluation.
• This approach has substantial potential for improving pavement management and maintenance strategies.
• Deep learning combined with physics-informed principles can effectively address challenges in detecting specific pavement defects.