Comparative study of pavement anomaly detection using detection models with rotated bounding boxes.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Civil Engineering

Background:

• Automated pavement anomaly detection is crucial for infrastructure maintenance.
• Existing models like YOLOv4-Tiny have limitations in accuracy for detecting pavement defects.
• Pavement anomalies such as cracks and potholes require precise identification for effective repair.

Purpose of the Study:

• To enhance the accuracy of automated pavement anomaly detection models.
• To introduce and evaluate a novel rotated rectangle labeling strategy.
• To compare the performance of different YOLOv4-based models with and without rotated bounding boxes.

Main Methods:

• Developed and compared YOLOv4-Tiny and YOLOv4-ResNet50 models.
• Introduced a rotated rectangle bounding box strategy for labeling pavement anomalies.
• Evaluated models on a dataset of 1,107 cracks and 691 potholes from Nantong, China.

Main Results:

• The baseline YOLOv4-Tiny model achieved a mean average precision (mAP) below 0.4.
• The YOLOv4-ResNet50-rotated model achieved a superior mAP of 0.742.
• Rotated bounding boxes improved the detection of inclined cracks and potholes compared to axis-aligned boxes.

Conclusions:

• Rotated rectangle labeling significantly enhances pavement anomaly detection accuracy.
• The YOLOv4-ResNet50-rotated model demonstrates state-of-the-art performance in detecting pavement defects.
• This approach provides a foundation for more accurate and robust infrastructure monitoring systems.