STRUCTURE AND MOTION WITHOUT CORRESPONDENCES

This thesis investigates a novel paradigm to estimate motion and epipolar geometry constraints from orthographic images of a three-dimensional scene. This new paradigm tackles the challenges that conventional correspondence-based estimation approaches are inevitably exposed to, since reliable correspondences are hard to obtain in practice. The estimation problem is solved directly from either image sets or unmatched sets of feature points, without the need for correspondences and without imposing additional constraints on the data. Initially, a technique is proposed that deduces 3D motion information directly from a set of planar stereo images. This method operates in the Fourier domain and is based on the Fourier slice theorem commonly used in computer tomography. The foundation of the underlying theory is in the deployment of an integral projection scheme as a pseudo-measurement model to describe orthographic projection. Three different estimation methods are experimentally evaluated. The most robust approach among them is based on a maximum likelihood rationale and has been specifically designed to cope with distortions in the input data. The two main limitations of the proposed technique are analysed both theoretically and experimentally. They consist of the restriction of the technique to planar 3D scene objects and the limitation of the estimation accuracy resulting from inevitable signal processing artifacts. Both these limitations are avoided when the method is extended to sets of sparse features. It is shown that the epipolar geometry of the stereo configuration and thus the fundamental matrix can be derived from the parameters of interest estimated by the proposed method. The behaviour of the corresponding algorithm is then thoroughly evaluated using both synthetic and real image data