PSI - Issue 84

Francesco Allegrezza et al. / Procedia Structural Integrity 84 (2026) 81–88

82

1. Introduction Structural Health Monitoring (SHM) is a fundamental tool for ensuring safety, reliability, and efficient management of infrastructures, e.g., Wenzel (2009), Farrar and Worden (2013), Webb et al. (2015), Chen (2018). Advances in this field are closely linked to the development of increasingly accurate measurement technologies capable of operating continuously and with minimal impact during the service life of the monitored structures. Among the various parameters of interest, displacement measurements play a central role, as they provide direct information on the overall behaviour of the structure and its response to external stresses. In the case of bridges, measuring deck displacement is particularly important both for characterizing slow movements, related for example to thermal, viscous or landslide deformations, e.g., Meoni et al. (2024), and for analysing the dynamic response induced by vehicular traffic or seismic events, e.g., Conte et al. (2008), Ubertini et al. (2013), Scozzese and Dall’Asta (2024), Scozzese et al. (2025). However, direct measurement of displacements presents considerable operational difficulties, especially in the absence of stationary reference points, as is typically the case at midspan. For this reason, in practical applications, dynamic monitoring of bridges is often performed using accelerometers, from which the displacement must be reconstructed by numerical integration, with consequent problems related to error accumulation and noise sensitivity. In this context, computer vision techniques are establishing themselves as a promising alternative to traditional measurement systems, e.g., Feng and Feng (2018), Dong and Catbas (2021), Zona (2021), not only in bridge monitoring but also in multistorey buildings, e.g., Gioiella et al. (2025). The analysis of video sequences allows information on movements to be extracted directly and without contact, achieving subpixel resolutions and acquisition frequencies compatible with the study of the dynamic response of structures. Additional advantages of this approach include ease of installation, the ability to monitor multiple points simultaneously from a single sensor (video camera), and relatively low hardware costs compared to conventional sensors. The vision-based system adopted in this work relies on the use of fiducial markers for automatic target identification and on area-based template matching algorithms supplemented by specific subpixel estimation strategies to obtain the displacement measurements, building on the developments presented in Micozzi et al. (2023, 2024, 2025). This configuration allows for robust definition of regions of interest and accurate tracking of multiple targets using a single camera. The main objective of this study is to evaluate the accuracy of two different subpixel estimation approaches, analysing their behaviour under controlled and real conditions. To this end, the algorithms were tested on synthetic videos with known displacements, on laboratory tests with controlled translations, and on a full-scale application involving a post-tensioned concrete bridge in service. The results show very good accuracy in the reconstruction of displacements and strong consistency with reference measurements, also highlighting how some systematic bias effects observed under ideal conditions tend to attenuate in real acquisitions. Overall, the work confirms the potential of computer vision techniques as reliable tools for structural monitoring of displacements of bridges. 2. Methodology The algorithms employed in this study for the displacement identification are based on template matching techniques, in which the position of the template is obtained by maximizing a similarity measure, and the displacement between different frames is computed as the difference between the previously estimated positions. Specifically, the chosen similarity measures are the Phase Correlation (PC) and the Zero-Normalized Cross Correlation (ZNCC). These methods grant accuracy only up to the pixel level, so to improve the displacement estimate two distinct sub-pixel algorithms were associated with each similarity measure. After the PC, the upsampling algorithm developed by Guizar-Sicairos et al. (2008) based on DFT-matrix multiplication is adopted, whereas a bidimensional quadratic function was fitted to the data provided by the ZNCC. In the first case, the accuracy level is directly controlled by the user through an upsampling factor; in this study this factor was set to 100, corresponding to a sub pixel resolution of 1/100 of a pixel. In the other case there is no explicit parameter that regulates the accuracy of the estimate: the position that indicates the best similarity is found by maximizing the interpolating function, and its accuracy depends on the quality of the approximated surface.