Feature matching for automated and reliable initialization in three-dimensional digital image correlation

In three-dimensional digital image correlation (3D-DIC), corresponding locations must be matched in a pair of images either across time or across cameras. The most effective approach is to optimize the correlation of the related intensities by using iterative algorithm (Newton–Raphson or Levenberg–Marquardt). However, the iterative optimization requires accurate initial guess of the unknown parameter. We hereby present an automated and reliable initialization method which utilizes image feature matching. Feature points are first detected in the images with high repeatability and each feature is characterized by a descriptor which is insensitive to common image transformations. The features are then matched across images based on the descriptor similarity and a geometric constraint. The deformation parameter of a point of interest is initially estimated from the affine transformation fitted to the matched features inside the subset area. Experimental results on simulations show that the proposed initialization is sufficiently accurate to enable correct convergence of the subsequent optimization, in the presence of rigid motion and heterogeneous deformation of the specimen and variation in camera viewpoint. The performances on real-world experiments also verify the accuracy and robustness of the method in both temporal matching and stereo matching in 3D-DIC.

► We utilize image feature points for automated and reliable initialization of 3D-DIC.
► Feature representation is discriminative and robust to image transformations.
► Features are robustly matched despite large image transformation.
► Deformation is estimated by affine transform fitted to local matched features.
► Initialization ensures correct convergence of iterative optimization.