Enhancing pavement health assessment: An attention-based approach for accurate crack detection, measurement, and mapping

Abstract

Vehicle-based crack detection can serve as a highly effective method for assessing pavement damage. This paper proposes an automated crack detection, measurement and mapping method based on GPS tagged images acquired from a camera installed on the license plate of a vehicle. An attention-based mechanism is added to the multiscale feature fusion RetinaNet convolution neural network to detect cracks. This enables the neural network to focus on key features at different scales enhancing the model’s detection capabilities. Adding the attention block effectively increases the model’s performance when compared with the original network. This is followed by a lightweight categorization algorithm, a high-precision edge detection algorithm, a crack width measurement algorithm, and a mapping algorithm. The results show that the crack detection model is a well-represented model that exhibits high performance on crack-induced pavements. The process of estimating crack width is validated by achieving a low mean average relative error. The crack width is further used to assign severity to the crack. The crack detection algorithm and the crack width estimation algorithm are tested using real-world data in order to assess their effectiveness. The results of the study suggest that by integrating crack detection and width estimation, a comprehensive engineering solution for crack monitoring can be achieved. This solution can be effectively applied to conduct pavement surveys, resulting in the creation of a GPS-based map that identifies various distresses. This framework provides a holistic approach to infrastructure maintenance, allowing improved transportation infrastructure management.