Application of Ai-based Deformation Extract Function from a Road Surface Video to a Road Pavement Condition Assessment System

In Japan, there is a concern that civil infrastructure will rapidly age in the near future. This study focuses on asphalt road surfaces, which are typically renovated every 10 years depending on the amount of traffic and roadbed properties. Existing MCI (Maintenance Control Index) measurement systems come at a high cost to local governments and are not efficient in allowing engineers to detect cracks and deficiencies. New road pavement assessment systems, as developed by our research group, are needed to ensure sustainable road maintenance and management. Pavement surface evaluation system development involves the use of a video camera and a 3D motion sensor, which can be used for simple and low-cost inspections. However, 3D motion sensors can only capture acceleration. Because of this, they can only be used to illustrate the roughness of the road surface, not to detect cracks. In this study, to utilize road surface video recorded while driving, we have developed a method of automatic extraction of deformations by an AI object detection function. This function specifically serves to extract cracks, joints, manholes, and repair marks detections from the surface video. However, in using this function, the accuracy for detecting cracks was less than 40% (Shiga et al. 2020). In this study, we aim to apply this method to detect deformations and suggest annotation rules for improving the accuracy of crack detection, as well as overall accuracy. To discuss the accuracy of detecting cracks and other deformities, cracks are divided into different types and deep learning is performed. In addition, we enlarged images of the cracks.

The results of this study show that the AI object detection function for cracks is made more accurate by utilizing annotation rules and making a learning data rule set divided by the crack type classification.