ScholarWorks@성균관대학교

초록

Reliable evaluation of prestressed steel strand tension is vital for structural performance, yet existing methods remain costly and dependent on periodic calibration. Vision-based alternatives offer non-contact solutions, but they typically rely on external cable imagery, infer tension indirectly, require specialized equipment, and are vulnerable to environmental variation. To overcome these challenges, this paper introduces a vision-based multimodal framework that directly analyzes internal strand imagery obtained at arbitrary time points using standard-performance cameras. The framework is designed as a sequential process in which the output of each module serves as the input for the following stage, where SAM2 provides robust segmentation without task-specific training, Co-Tracker v3 captures dense pixel trajectories across frames, and a Lightweight Dense Convolutional Network (LDC) detects line segments that enable lay-angle estimation. These outputs are subsequently integrated into a mechanics-based formulation to compute strand tension, yielding a tightly connected pipeline from raw images to quantitative force evaluation. Experimental validation was conducted with a custom-built apparatus replicating embedded strand conditions, where cyclic loading–unloading tests achieved a mean absolute error of 3.16 % compared with load cell data. Beyond numerical accuracy, the framework avoids calibration, lowers hardware costs, and improves robustness by leveraging internal strand imagery rather than external cable views. In addition, its systematic sequential design provides a pathway toward automation, enabling a continuous workflow from image acquisition to tension estimation without manual intervention. The results highlight an end-to-end methodology that enables practical and automated monitoring of prestressed structures.

키워드