Rapid monitoring of structural deformation based on unsupervised segmentation model

Abstract

This paper presents a novel approach for rapid structural deformation monitoring using an enhanced segmentation model, SAM-DM (Segment Anything Model-Deformation Monitoring). The method enhances the SAM model by integrating an innovative point prompt generation technique, image fusion, and connected domain analysis, enabling structural deformation monitoring with just a single manual click on the target area in the initial video frame. Specifically, this manual click generates coordinate information for the target area, serving as a point prompt for the SAM-DM model. This process results in the creation of a high-quality binary mask, which is then fused with the original image to isolate only the target area. Subsequently, connected domain analysis is employed to automatically process the mask, extracting the pixel centroid coordinates of the target area, which are then used as prompts for segmenting subsequent frames. This approach allows continuous tracking of changes in the target area's coordinates, thereby facilitating rapid structural deformation monitoring. Experimental results demonstrate that this method can accurately monitor deformation across all target points and adapt to varying lighting conditions and noise levels. Applied to monitor the bending deformation of reinforced concrete beams and the in-plane deformation of concrete slabs, the method achieves results closely aligned with actual measurements, yielding relative errors of 0.95 % and 0.48 %, respectively. Additionally, in monitoring the vibrational deformation of steel frames, the primary frequency of the curve precisely matches measured values, confirming that the SAM-DM-based method delivers both high accuracy and strong robustness.