Geometric Principles of Multiple Visual Sensors

186 p.Thesis (Ph.D.)--University of Illinois at Urbana-Champaign, 2004.In this dissertation I study both the theoretical and practical issues in building a large scale system with distributed visual sensors. I have developed a unified algebraic approach for studying the structure-from-motion problem, which is a set of rank conditions on the multiple-view matrices of image features. This approach can be applied to characterize the algebraic conditions governing the multiple views of any type of image features and any kind of incidence relations. A set of linear iterative 3-D reconstruction algorithms have been developed based on this approach. It has also been generalized to the case of dynamical scenes and 3-D reconstruction from a single image of symmetric objects. In addition to the theory, I have studied three important practical issues: large-baseline feature matching, motion segmentation, and visual sensor synchronization. For feature matching, I developed an algorithm based on a graph theoretical approach, which uses symmetry as a high-level feature to perform matching for arbitrarily large baselines. I have designed a robust GPCA algorithm that can segment motions of different types. Finally, I have presented a visual sensor synchronization scheme purely from the visual input for arbitrarily moving cameras.U of I OnlyRestricted to the U of I community idenfinitely during batch ingest of legacy ETD