A Simplified Gaussian Approach for Asphalt Crack Detection based on Deep Learning and RGB images

Abstract. Monitoring pavement condition is a crucial aspect for pavement maintenance management systems (PMMS). There are several pavement characteristics that affect the pavement condition, Crack distress is a highly representative type of pavement distress and often serves as an early indicator of more extensive pavement issues. Cracks impact both the operational efficiency and safety of road pavements and significantly influence maintenance decisions. We propose a workflow to detect cracks using YOLOv9 deep learning algorithm combined with statistical analysis through principal component (PCA) and Gaussian distribution. The proposed workflow includes camera calibration to address the metric issues in vision-based crack detection methods, utilizing Zhang's calibration method to compute the camera's internal and external parameters. To validate the proposed framework, three different datasets were acquired. Laser Crack Measurement System (LCMS) was used as a ground truth data for further verification the proposed method. Experimental results demonstrate that the proposed method achieves millimeter-level accuracy (std= ±1.0mm) compared to LCMS. This indicates the method's potential applicability for asphalt road crack segmentation and crack width estimation.