A Review of Pavement Crack Detection Based on YOLO

Abstract

As transportation infrastructure expands, pavement distresses have become more frequent. Cracks, as the most common early damage, directly affect road safety and service life. Traditional manual inspections are time-consuming and highly subjective, while conventional image-processing methods lack robustness under varying illumination, noise, and complex textures. Deep learning, especially the YOLO family, offers a feasible solution for large-scale, fast, and automated crack detection thanks to its real-time performance and end-to-end detection capability. Improving detection accuracy under complex conditions, enhancing small-target recognition, and boosting model generalization have become key issues in current road maintenance and intelligent inspection. With the continuous expansion of road infrastructure scale, the operational safety and service performance of pavements have become increasingly critical to transportation systems. Pavement cracks are one of the most common and representative forms of structural distress, and their timely and accurate detection is of great significance for pavement maintenance decision-making and lifecycle management. Traditional manual inspection and classical image processing–based methods suffer from limitations such as low efficiency, strong subjectivity, and poor robustness under complex environmental conditions. In recent years, deep learning–based object detection methods have shown remarkable advantages in feature representation and detection accuracy, among which the You Only Look Once (YOLO) series algorithms have attracted extensive attention due to their real-time performance and end-to-end detection framework. This paper systematically reviews the research progress of pavement crack detection based on YOLO algorithms. First, the basic principles of pavement crack detection and the evolution of YOLO models are introduced, including their network structures, detection mechanisms, and performance characteristics across different versions. Then, existing studies applying YOLO models to pavement crack detection are comprehensively summarized from aspects such as dataset construction, annotation strategies, model training schemes, and parameter optimization methods. Typical research results are comparatively analyzed in terms of detection accuracy, speed, and adaptability under complex backgrounds, and the advantages and limitations of different YOLO variants are discussed. Furthermore, this paper compares YOLO-based methods with other deep learning detection and segmentation approaches, highlighting their relative strengths in engineering applications. Finally, current challenges, including small crack detection, data imbalance, environmental interference, and model generalization, are analyzed, and future research directions such as lightweight model design, multi-scale feature enhancement, and intelligent inspection system integration are proposed. This review aims to provide a systematic reference for researchers and engineers engaged in pavement distress detection and to promote the practical application of YOLO-based crack detection technologies.