A Hybrid Perspective of Vision-Based Methods for Estimating Structural Displacements Based on Mask Region-Based Convolutional Neural Networks

Abstract

Vision-based methods have shown great potential in vibration-based structural health monitoring (SHM), which can be classified as target-based and target-free methods. However, target-based methods cannot achieve subpixel accuracy, and target-free methods are sensitive to environmental effects. To this end, this paper proposed a hybrid perspective of vision-based methods for estimating structural displacements, based on Mask region-based convolutional neural networks (Mask R-CNNs). In proposed methods, Mask R-CNN is used to first locate the target region and then target-free vision-based methods are used to estimate structural displacements from the located target. The performances of proposed methods were validated in a shaking table test of a cold formed steel (CFS) wall system. It can be seen that Mask R-CNN can significantly improve the accuracy of feature point matching results of the target-free method. The comparisons of estimated structural displacements using proposed methods are conducted and detailed into accuracy, stability, and computational burden, to guide the selection of the proper proposed method for the specific problem in vibration-based SHM. Proposed methods can also achieve even 1/15 pixel-level accuracy. Moreover, different image denoising methods in different lighting conditions are compared.