Yi Yan, Xu Yang, Wenbo Liu, Hewei Zhao, Hainian Wang
The precise quantification of asphalt pavement cracks is essential for effective road maintenance and condition assessment. To address the challenge that traditional detection methods struggle to automatically determine the longitudinal profile dimensions of concealed cracks, this study proposes a keypoint detection model named KP-CrackNet for automatic measurement of the vertical start and end points of cracks. Based on finite-difference time-domain (FDTD) forward modelling and laboratory experiments, the typical characteristics of ground penetrating radar (GPR) images of concealed crack profiles were analysed, and representation formulas for crack length and inclination angle were established. An improved YOLOv11-Pose model was developed by integrating a global attention mechanism (GAM) and an OKSClip loss module, enabling the automatic localisation of endpoint keypoints and subsequent estimation of crack inclination and length. Experimental results show that the F1 scores of the n and l models reached 0.718 and 0.712, representing improvements of 6.21% and 5.17% over the baseline model. The automatic measurement errors for length and angle were 6.40% and 4.86%, respectively. This method provides reliable technical support for non-destructive, rapid, and quantitative crack detection and for intelligent pavement maintenance decision-making.