Development of Pavement Distress Detection Model Utilizing SOTA Deep Learning Algorithm “YOLOv5”

One of the most critical issues in pavement asset management is evaluating the performance of the roads and highways. This crucial task is currently handled by regular manual inspection in many countries, which is inaccurate and sometimes dangerous. However, this inspection is processed automatically utilizing specifically designed vehicles in some developed countries; many municipalities and road agencies worldwide are still using manual methods due to the high expenses of purchasing and maintaining specific vehicles. Due to the recent advancements in computer vision, researchers and scholars use deep learning technology to enhance road inspection. Some high-tech infrastructure cities benefit from this technology to handle various infrastructural issues. Roads and pavements are no exception in this area. Currently, some intelligent cities are using deep learning technology to evaluate road performance. Pavement distress detection is one of the critical issues in this field. Many researchers worldwide use deep learning technics, expressly object detection algorithms, to automate road performance evaluation. This study aims to provide a robust and reliable model for detecting and classifying several types of pavement distresses with high accuracy. Road agencies and municipalities could use this model to collect data on road sections conveniently and affordably. In this case, authorities could monitor the pavement condition in short intervals and make appropriate decisions on maintenance and rehabilitation strategies and methods which could result in maintaining the performance of the pavement an acceptable quality by lower costs. This study developed a model to detect and classify pavement distresses on the surface of the road utilizing state of the art deep learning algorithm (YOLOv5) as well as most recent prominent optimization strategies such as data augmentation and hyperparameter tuning to propose an accurate, robust, and reliable model. 628 topdown view pavement images used in this study were captured in several cities in the U.S., including various distress types such as alligator cracking, longitudinal cracking, transverse cracking, block cracking, patching, sealing, and manhole. The performance of the proposed model is evaluated based on several criteria. The model's accuracy reached 0.95, 0.92, and 0.93 in precision, recall, and F1 score, respectively.