UAV-Based Data Acquisition Strategy for 0.2 mm Crack Detection in Bridge Inspection

Abstract

During the service life of bridges, early-stage defects such as cracks are inevitable. Among them, fine cracks with a width of approximately 0.2 mm are critical control indicators in structural safety assessment and durability evaluation, and bridge inspection codes explicitly require their detection and quantification. However, bridge towers and other tall components contain extensive areas that are difficult to access through manual inspection. Although non-contact inspection based on unmanned aerial vehicles (UAV) offers significant advantages in safety and efficiency, reliably identifying 0.2 mm cracks at practical inspection distances remains a key engineering challenge. To address this issue, this study focuses on developing a data acquisition strategy for UAV inspection that is aligned with the 0.2 mm crack threshold specified in bridge inspection codes. Particular attention is given to the influence of imaging distance and illumination conditions on crack recognizability. An already damaged beam was selected as the experimental subject, and a DJI Matrice 4T aircraft was employed to collect crack images under systematically controlled distance conditions. The flight path was arranged perpendicular to the beam surface, with constant gimbal orientation and imaging angle to minimize pose-related variability. Image acquisition started at 2.5 m from the target surface and increased in 0.5 m intervals, reaching a maximum distance of 45.0 m. A total of 280 crack images were collected under clear afternoon conditions. Target cracks with widths of approximately 0.2 mm were identified using a crack width comparison gauge. Pixel-level annotations were performed using an open-source image annotation tool, and crack recognition was evaluated through a semantic segmentation task implemented with a standard version of the You Only Look Once 11(YOLO11) model, without architectural modification. This choice was made to avoid dependence on model-specific enhancements and to better ensure the general applicability of the proposed data acquisition strategy across commonly used detection frameworks. Illumination effects were preliminarily analyzed by applying synthetic brightness adjustments to the collected images. Under these controlled adjustments, the segmentation recall remained above 0.80 and the mean Intersection over Union (mIoU) remained above 0.65, suggesting limited sensitivity to moderate brightness variation. However, this conclusion is restricted to simulated brightness changes and does not account for complex real-world lighting factors such as shadows, glare, and non-uniform illumination. The influence of imaging distance was systematically evaluated. Results indicate that 0.2 mm cracks can be reliably recognized when the imaging distance does not exceed 25.5 m. Based on camera imaging principles, the corresponding ground sampling distance was calculated to be approximately 1.29 mm per pixel. Given known camera parameters, the maximum permissible imaging distance for satisfying the 0.2 mm detection requirement can be inversely derived. The findings provide practical guidance for planning inspection flights and selecting appropriate imaging distances in bridge crack detection tasks subject to code-specified crack width limits.