End‐to‐end instance segmentation framework for automated pavement distress assessment using unmanned aerial vehicle‐derived digital orthophoto maps

This study presents a novel framework for automated detection and segmentation of pavement distresses using high-resolution digital orthophoto maps captured by unmanned aerial vehicles (UAVs). While recent UAV-based research often relies on manual measurements from rasterized digital models, only a few studies have employed automated methods for analyzing pavement distress, primarily using raw aerial image tiles. Developing a distress detection system specifically suited for UAV-based, scale-variant orthomaps is complex but offers significant advantages. This study developed a comprehensive dataset from UAV-derived orthophotos and utilized geographic information system tools to create precise vector annotations of various pavement distress types. These annotations were transformed into standard object detection formats to train state-of-the-art computer vision models. The study evaluated both one-stage (YOLOv7, YOLOv8 with YOLACT [You Only Look At CoefficienTs] segmentation head) and two-stage (Mask R-CNN) instance segmentation networks under various configurations. The YOLOv7 model with low hyperparameter settings achieved the highest performance, attaining mean average precision scores of 0.70 for box detection and 0.67 for mask segmentation. Inferencing strategies, such as strided inference with different tile sizes, were employed to optimize detection across varying scales. The detection results from the YOLOv7 model demonstrated a satisfactory agreement with manual measurements, highlighting the framework's potential for project-level pavement assessment.