Computational intelligence for road pavement condition assessment: a deep learning perspective

Abstract Pavement defects such as cracks and potholes compromise road safety and demand timely maintenance. Traditional manual inspection is slow and exposes workers to safety risks, whereas automated systems offer a promising alternative. This survey provides a comprehensive review of deep learning methods for road condition assessment. We first examine 2D image-based approaches, tracing their evolution from convolutional neural networks (CNNs) to Transformers. Although these methods are widely adopted, they remain sensitive to lighting conditions and cannot directly capture physical properties such as defect depth. To address these limitations, we review 3D sensing and subsurface diagnostic techniques, which provide essential geometric information for severity assessment. The primary focus of this paper is on evaluation: we summarize key public datasets and evaluation metrics and analyze the persistent gap between algorithmic performance and the practical needs of engineering, emphasizing the importance of assessing the actual utility of models in the field. Finally, we discuss several key challenges and promising research avenues, arguing that future work should prioritize model robustness, reliability, and the integration of these systems into real-world maintenance workflows.