PSI - Issue 84

Giorgia Ghirelli et al. / Procedia Structural Integrity 84 (2026) 1047–1054

1051

5. Results This section presents the outcomes of the experimental monitoring campaign and the numerical analyses performed on the selected span of the bridge. The discussion begins with the vision-based measurements, focusing on the performance of different feature identification and tracking strategies used to reconstruct the dynamic displacement of the checkerboard targets. Subsequently, the displacement time histories extracted from the videos are compared with the corresponding predictions obtained from the Finite Element model, allowing for a comprehensive evaluation of the proposed methodology. 5.1. Experimental vision-based monitoring The vision-based analysis aims to reconstruct the motion of the checkerboard corners, which represent highly distinctive features in the acquired images. To evaluate the robustness and accuracy of the vision-based workflow, alternative processing strategies are considered and compared. The processing pipeline relies on two main phases: (i) corner detection and (ii) temporal tracking of the detected features across frames. All strategies share a common initial step: the identification of checkerboard corners. Although several corner detection algorithms exist in the literature, this study adopts the methodology proposed by Geiger et al. (2020). It is a robust and fully automatic approach for detecting checkerboard corners – including sub-pixel refinement – from a single frame, even under varying imaging conditions. The resulting feature set consists of well-defined and highly trackable points, allowing for sub-pixel accuracy when combined with appropriate image-processing routines and camera calibration. Two approaches are considered for tracking the checkerboard corners over time. The first, called frame-by-frame identification, applies the corner-detection procedure independently to each video frame. The second, called temporal tracking, updates the position of each feature of the first frame based on local optical-flow estimation. To this purpose, the methodology introduced by Lucas and Kanade (1981) and further developed by Tomasi and Kanade (1991) is here applied. The results show that the temporal-tracking approach offers significantly lower computational cost. However, it is more sensitive to temporary loss of features, which may occur under rapid lighting variations, partial occlusions, or motion blur. When a feature is lost, the tracker may drift or fail to update the position, thereby introducing local inconsistencies in the displacement time history. Conversely, the frame-by-frame re-identification strategy ensures a more robust and stable reconstruction of corner trajectories. Since the detection is repeated independently at each frame, the method does not accumulate errors over time and is less affected by changes in environmental or illumination conditions. This results in more reliable displacement measurements, particularly in cases where the structural motion is small and must be captured at sub-pixel precision. The trade-off, however, is a higher computational demand due to the repeated execution of the detection algorithm. To balance the advantages of both methods, the present research introduces a hybrid methodology. The hybrid strategy tracks the features across consecutive frames to exploit the efficiency of optical flow but performs a full corner re-identification at regular intervals. In this way, the algorithm periodically “resets” the tracked positions, eliminating drift and correcting potential mismatches before they accumulate. The frequency of re-identification is selected based on the frame rate and the expected motion amplitude, ensuring that the checkerboard geometry remains consistently recognized throughout the sequence. This approach provides stable and accurate displacement time histories while maintaining manageable processing times. The different approaches are tested and compared based on the displacement time histories. To this end, the results corresponding to a target mounted on a transverse beam and captured by the camera positioned beneath the bridge (see Fig. 2b) are presented in Fig. 3. The averaged displacement time histories obtained from the three algorithms, namely frame-by-frame re-identification, feature temporal tracking, and the hybrid approach, exhibit nearly identical trends, peak values, and overall displacement amplitudes. The time series clearly reveal the passage of the train approximately 40 seconds after the start of data acquisition. This is initially represented by a prominent peak, corresponding to the leading locomotive, followed by smaller displacements as each carriage moves through. The passage of the final tail carriage is also clearly visible, generating a displacement smaller than that of the leading locomotive but larger than that of the intermediate carriages. Additionally, a negative displacement is observed,