Feature‐assisted stereo correlation

Abstract

Stereo correlation in digital image correlation (DIC) involves an optimisation problem that is sensitive to initial guess. In practice, this problem is circumvented by manually selecting a pair of points in the two stereo images that guarantees convergence and provides stereo mapping parameter estimates that are used as initial guesses at neighbouring subsets. However, such an approach is not always feasible, especially in the presence of substantial perspective distortions, for example, due to large stereo angles or complexities in specimen geometry. Therefore, it is desirable to provide high-quality independent initial estimates over the entire region of interest. Recently, SIFT has been used for this purpose, but it fails when perspective distortions are severe. In this work, we investigate seven other feature-based matching techniques to address this gap.Among these, DeepFlow algorithm provides the highest quality and most spatially uniform initial estimates. Further, we use DeepFlow estimates as initial guesses in a conventional stereo optimisation to compute geometry measures of a specimen in DIC challenge dataset. These geometry measures show excellent agreement with ground truth, further supporting the choice of DeepFlow in stereo correlation.

Three-dimensional shape and deformation measurement on complex structure parts

“…This is because sub-pixel iteration algorithm such as IC-GN algorithm needs a fine initial guess to guarantee the convergence. The introduced SIFT algorithm can be used for this purpose, but it does not perform well when there is severe perspective distortion since the corresponding features appear significantly different in the images of two perspectives 49 . Furthermore, if the tested specimen has complex structures such as deep hole, sharp edges, thin wall, high curvature regions or concave-convex structure, it becomes more challenging for stereo-DIC.…”

On improving the accuracy of self-calibrated stereo digital image correlation system

“…The distortions in the left and right images are removed by applying the distortion-induced corrections at each pixel location. The pose of the right camera with respect to the left camera defined by translation vector t v and rotation vector r v is determined by drawing correspondence across each of the 10 undistorted stereo pairs of images using our recent feature-assisted stereo correlation method [37], in which DeepFlow [38] initial estimates of the homography shape function parameters are taken to sub-pixel accuracy using the conventional DIC.…”

“…Six corners in the gauge area (points A to F shown in figure 5) are manually selected. Stereo correspondence is drawn at each of these six points again by using feature-assisted stereo correlation [37]. The points A to F are then triangulated and the lengths di , i = 1 ••• 4 of the line segments CA, CB, FD, and FE are calculated.…”

“…This idea of conducting a pilot experiment is adopted in this work for choosing not only the best aperture size, but also the best settings for self calibration. to the specimen a displacement of about 2 mm along e 3 shown in figure 2(a), (c) the calibration parameters are estimated by following the self-calibration procedure detailed in section 3, (d) the RoI is selected such that it spans the entire specimen gauge area, (e) the epipolar error 3 of the first stereo image pair r 1 and that of the second one r 2 are computed by using featureassisted stereo correlation [37], (f) the change in epipolar error r 1 − r 2 is calculated at each point of interest, and the standard deviation ||r 1 − r 2 || std is finally calculated 4 .…”