PSI - Issue 17

Francisco Barros et al. / Procedia Structural Integrity 17 (2019) 986–991 Author name / Structural Integrity Procedia 00 (2019) 000 – 000

989

4

intrinsic parameters were imposed as constant. The 3D positions of features from objects in the background and the positions assumed by the camera throughout the acquisition can be seen in Fig. 3.

Fig. 3. Three-dimensional representation of the mapped points (in grey) and the camera positions (in red) obtained from the structure from motion algorithm. The speckled plate corresponds to the region with high point density next to the camera locations.

The fact that a long video of the reference state was used is due to the fact that a larger quantity of images that show features in the background from different positions and camera orientations will decrease the uncertainty of the SfM results. The speckle pattern itself also provides robust detectable features. From the video of the loaded state, only the two necessary images were used, since the area to blur must be selected manually for each image. The SfM algorithm provides the intrinsic parameters and the camera orientations for all four images that were used for DIC, but their relative positions, and consequently the 3D positions of points obtained through triangulation, can only be obtained up to a scale factor. Therefore, the positions of the two corners of the checkerboard placed beside the plate, known to be 10 mm apart, were triangulated using the image pair from the reference state, according to the parameters from the SfM algorithm. The ratio between the real distance between them and the distance obtained in this triangulation is the aforementioned scale factor. The positions of the cameras obtained through SfM were multiplied by this scale factor in order to make the calibration dimensionally correct. Commercial DIC code was used to perform the subset matching algorithm itself. After the images were corrected for radial distortion, an area of interest was defined in one of the reference images and subsets were matched across all four images with a subset size of 29 px using a 2D-DIC program. Stereo DIC software is not necessary, since its advantageous functionalities are not prepared for the unconventional acquisition and calibration methods. 3D point triangulation and post-processing of obtained point clouds are, therefore, performed separately, leaving only the speckle pattern correspondences between images to the DIC software. After subset correspondences were found, the point clouds for the reference state and the loaded state were triangulated using the camera parameters obtained and the optimal triangulation method outlined by Hartley and Zisserman (2003). The point clouds were rotated such that the plane defined by the initial shape of the plate was parallel to the x-y plane, and the displacement in the z direction, i.e. perpendicular to the plate, was evaluated. 2.3. DIC and point clouds