A Novel Deep Learning Approach for Enhanced Roadway Pothole Detection Using YOLOv8 Instance Segmentation Algorithms

Potholes pose a major risk to roadway safety and vehicle durability, necessitating timely detection and repair. With advancements in artificial intelligence, deep learning has become crucial for automating pothole detection and segmentation. Previous studies using CNN and Haar cascade methods have achieved accuracies up to 98.2%. This paper presents a novel approach leveraging the YOLOv8 architecture for instance segmentation, enhancing detection accuracy by capturing contextual and spatial relationships. The process involves data collection, annotation, preprocessing, and model training using datasets from Roboflow Universe and Kaggle. The model's performance is assessed through sensitivity, precision, recall, F1 score, and mean average precision (mAP). Experimental results indicate a significant improvement, achieving a 99.2% mAP in pothole detection and segmentation. These findings highlight the potential of YOLOv8 in advancing automated road maintenance, ensuring safer and more efficient transportation systems.