PSI - Issue 84

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

83

The adopted computer-vison algorithms can monitor simultaneously multiple targets with a single camera allowing for displacement time histories of several points and relative displacement calculations. This is possible by defining dedicated Regions of Interest (ROIs) within the frame to constrain the template search, since during synthetic and laboratory test the displacement history is known in advance while in field tests the expected displacement range is generally predictable. To automate target identification and improve process repeatability, the system relies on Apriltags, originally proposed by Olson (2011). These fiducial markers consist of a square grid of black and white cells bordered by a black outer edge. Several families of these markers are available, each characterized by a specific internal grid dimension. In this study, targets belonging to the “36h11” family were employed, which have a 6x6 internal grid. AprilTags are associated with a detection software that determines position and orientation of each tag relative to the camera. The integration of this feature allows for precise identification of the templates to be monitored, the definition of the corresponding ROIs and the establishment of their dimensions in pixel units, thus, reducing variability due to manual intervention. 3. Experimental results 3.1. Case study 1: Synthetic videos The proposed algorithms were initially evaluated using a set of computer-generated grayscale videos, with the aim of quantifying the accuracy of the subpixel estimates under controlled conditions and with precisely known displacements. Imposing subpixel movements is not a trivial operation and requires a specific treatment, as a non integer translation cannot be achieved as a simple shift of the target on the pixel grid. Under such conditions, the information content originally associated with a single pixel would be redistributed across multiple positions, generating a configuration that cannot be represented correctly while maintaining the same resolution. This behaviour reflects what happens in real acquisitions, when the template does not align perfectly with the discretization imposed by the sensor. To simulate this scenario, an intermediate position was obtained by bilinear interpolation between the two frames with the closest integer positions. The trajectory assigned to the target was defined as a uniform motion along the line passing through the centre of the image and inclined by 30° with respect to the horizontal. This choice allows the performance of the algorithms to be evaluated simultaneously on both horizontal and vertical components and to verify the effectiveness of the subpixel interpolation. To ensure consistency with laboratory and field tests, the synthetic videos reproduced the same configuration of AprilTag targets used in real acquisitions. The generated videos have a resolution of 300  300 pixels, chosen because it is comparable to the typical size of the ROIs used during the analysis of experimental data; each target cell was modelled as a 10  10-pixel square region, resulting in an internal matrix of 60  60 pixels. All videos were generated in grayscale 8-bit single-channel format, i.e., each pixel assumes an integer number between 0 and 255. Two kinds of synthetic videos were generated (Fig.1): one characterized by a blurring effect (Fig.1a) that simulates real-world conditions were perfectly sharps edges are never completely obtained, introduced by convolving each frame with a Gaussian kernel (σ = 3) and one (Fig.1b) without any kind of image degradation and perfectly sharp edges of the target.

a

b

Fig. 1. Synthetic reproduction of AprilTag with blur (a) and without blur (b).