@Article{sdhm.2026.083682,
AUTHOR = {Zongqi Xiong, Yu Li, Xinxin Cao},
TITLE = {High-Fidelity Characterization and Physical Quantification of Bridge Pier Cracks Based on Oriented Object Detection},
JOURNAL = {Structural Durability \& Health Monitoring},
VOLUME = {},
YEAR = {},
NUMBER = {},
PAGES = {{pages}},
URL = {http://www.techscience.com/sdhm/online/detail/27230},
ISSN = {1930-2991},
ABSTRACT = {Accurate detection and quantitative characterization of bridge pier cracks are essential for UAV-assisted structural health monitoring, but conventional horizontal bounding box detectors often suffer from geometric mismatch, background redundancy, and missed detection of slender cracks in dense regions. To address these limitations, this study proposes an oriented object detection-based framework for high-fidelity crack representation and physical quantification. A dedicated PierCrack-OBB dataset was first constructed from real bridge inspection images and annotated using multi-segment oriented bounding boxes to better represent locally curved and high-aspect-ratio cracks. On this basis, the standard YOLOv8-OBB framework was improved by incorporating DCNv3 for adaptive crack feature extraction, a channel-spatial attention module for background suppression, and Gaussian probability loss for stable rotated-box regression. In addition, the minor-axis and inclination parameters of the detected oriented boxes were combined with the ground sample distance of UAV images to estimate physical crack width and describe crack orientation characteristics. Experimental results show that the proposed method increased mAP@0.5 from 88.6% for standard YOLOv8-OBB to 91.6%, while achieving 92.8% precision and 90.1% recall. For crack width estimation, the mean absolute error remained below 0.18 mm. These results demonstrate that the proposed method improves both the geometric fidelity of crack detection and the engineering interpretability of detection outputs. The main novelty of this study lies in integrating multi-segment OBB-based crack representation, improved oriented detection, and GSD-based physical quantification into a unified framework, thereby extending crack detection from image-level localization to measurable width estimation and orientation characterization.},
DOI = {10.32604/sdhm.2026.083682}
}