Reconstruction Of Three-dimensional Motion Of Deformable Surfaces Using Stereovision And Its Applications

The surface of an object can undergo complex three-dimensional (3D) mo-tion due to applied external forces and temperature change. Such phenomenon prevalently exists in nature and it has drawn the attention of scientists of multiple disciplines. Quantitatively analysis of surface motion has many applications in fields including computer science, material science, experimental mechanics, etc. The most essential task during the analysis would be accurate measurement of the3D surface motion. The recent advance of computer science allows us to utilize video cameras and computer vision technology to acquire the motion information of a deforming surface.The3D motion of a dynamic surface can be represented by the3D motion trajectories of a dense set of sample points on the surface. The time-varying3D positions of each sample point can be reconstructed after finding its corresponding locations in the images recorded at different viewpoints and time. However, image matching across viewpoint and across time is an extremely challenging task. The major difficulty is that the location, shape and intensity of the local image around each sample point could change drastically due to viewpoint variation, surface deformation, illumination change and self-occlusion. To resolve this problem, we proposed in this dissertation two methods for establishing dense correspondences between images in an accurate, robust and automated manner.The feature-assisted coarse-to-fine approach proposed in this dissertation uti-lizes the fact that the detection and descriptor of image feature points are in-sensitive to common image transformations. First, the feature points detected at different time and different viewpoints can be correctly matched with higher prob-ability by utilizing temporal and stereo constraints. Based on the resultant sparse matches, the correspondence of each sample point is initially estimated by fitting an affine transformation to the matches in its vicinity and subsequently refined by using iterative optimization. By examining the similarity between the reference image recorded before surface deformation and each target image in the sequences, potential self-occlusion and subsequent resumption of the sample points can be au-tomatically detected and error accumulation during the tracking in a long period of time can be totally avoided.The proposed optimized parameter transfer makes use of the relationship be-tween the sample points and utilizes the correspondence of a previously analyzed point to estimate the correspondences in its vicinity. Unlike the traditional pa-rameter transfer which directly copies parameters among points, a propagation function is presented to quantitatively model the potential discrepancy between the motion parameters of the adjacent points. The proposed method produces more precise initialization and therefore greatly improves the accuracy and ef-ficiency of subsequent iterative optimization of the motion parameters. It also allows more flexible choice of sampling interval and rapid motion change across the sample points. In addition, the point with highest matching reliability is se-lected in each transfer, which significantly alleviates the error propagation caused by transferring from incorrect results.By using the above methods, the complex3D deformation of fabric and hu-man face can be successfully reconstructed. In addition, we have.applied the feature-assisted approach and the optimized parameter transfer to deformation measurement in experimental mechanics. The widely used approach, digital im-age correlation, can be greatly improved by the proposed methods in measurement precision, robustness, computational efficiency and automation. By embedding the proposed method, digital image correlation is able to achieve desirable results in measuring very large deformations.