Machine Learning-Based Road Damage Detection: A Comprehensive Review and Comparative Analysis

Abstract

Finding road damage is essential to maintaining transportation safety, cutting maintenance expenses, and extending the life of infrastructure. Manual surveys and sensor-based systems, two traditional approaches to road damage assessment, are often labor-intensive, resource- intensive, and unsuitable for widespread use, especially in settings with limited resources. In this thesis, machine learning-based methods for detecting road damage are thoroughly reviewed and compared, with an emphasis on utilizing cutting-edge deep learning models, specifically YOLOv8 and its variations (YOLOv8n-seg, YOLOv8s-seg, YOLOv8m-seg, YOLOv8l-seg, and YOLOv8x-seg). Using picture datasets gathered from various contexts, the research investigates how well these models perform in identifying and categorizing different kinds of road damage, including cracks, potholes, and surface wear. To identify the best methods for practical applications, important factors including accuracy, computing efficiency, scalability, and cost-effectiveness are assessed. The findings show howsophisticated deep learning methods may be used to identify road degradation in a reliable, effective, and scalable manner, facilitating preventative maintenance plans and improving traffic safety. Limitations including dataset bias, processing needs, and environmental unpredictability are noted, despite the fact that the results show notable advancements in automated road damage identification. These difficulties highlight the need for more flexible and approachable solutions in future studies. By providing insights for scholars, decision- makers, and practitioners seeking to update road maintenance systems, this study advances the area of machine learning-based infrastructure management.