Application of Deep Learning on Structure Displacement Measurement and Accuracy Analysis

Abstract

Displacement serves as a crucial indicator in structural health monitoring within civil engineering, with computer vision being a commonly employed tool due to its cost-effectiveness and convenience. This paper introduces a methodology for tracking feature points computed via computer vision techniques SuperPoint and SuperGlue. Two test scenarios are conducted: a bridge loading test characterized by a low uniform stress state, and a tensile strength test marked by a high stress state. However, a notable drawback of this method is the potential misalignment between the survey point (the point under measurement) and the feature point, leading the errors. To address this issue, a fundamental principle for selecting feature points is presented. In the bridge loading test, the displacement of the survey point is recorded, and the influence of misalignment between survey point and feature point on accuracy is analyzed under low-stress uniform conditions. The findings demonstrate the feasibility of this method for field measurements in civil engineering. Subsequently, a tensile strength test is conducted to assess the influence of different feature point selection criteria on accuracy under high-stress conditions. Concurrently, the DIC software Ncorr is applied to analyze the displacement field of feature points near the survey point. The results highlight the significant influence of selecting feature points within regions of high stress concentration on measurement accuracy.