Fully automated pavement crack quantification and PCI estimation via a digital twin-ready vision-based deep learning pipeline

Effective maintenance of transportation infrastructure relies on the accurate and timely assessment of pavement conditions. Traditional manual inspections, however, are labor-intensive, subjective, and hazardous. While deep learning has advanced automated distress detection, a critical gap remains in translating computer vision outputs into standardized engineering metrics suitable for Digital Twins. This study proposes an automated engineering integration framework for pavement distress assessment and quantification. The methodology integrates YOLOv12, a state-of-the-art object detection model, with DeepLabV3+ for high-fidelity semantic segmentation. The system was trained and validated on a comprehensive multi-source dataset comprising public benchmarks (Crack500) and crowdsourced field data. A novel post-processing algorithm was developed to convert segmentation masks into geometric parameters (length, width, area) and automatically calculate the Surface Crack Index (SCI) and estimate the Pavement Condition Index (PCI)—based exclusively on longitudinal, transverse, and alligator cracking—in strict accordance with the deductive logic of ASTM D6433 standards . Experimental results demonstrate that the YOLOv12 model achieves a Mean Average Precision (mAP @ 0.5) of 0.861, while the DeepLabV3+ model attains an Intersection over Union (IoU) of 0.94. The automated assessment module showed a relative error of less than 5% when compared to manual field measurements. This framework effectively bridges the gap between raw visual data and actionable maintenance intelligence, providing the structured semantic data outputs necessary to feed dynamic, data-driven Digital Twin environments.