The development of flexible camera setup for tabletop augmented reality application

The prospective of computer vision technology in tabletop application is usually rendered unex-plored due to several critical points i.e. the availability of a robust algorithm to interpret information from images and the scalability problem. The traditional pin-hole mathematical model for camera representation as a baseline for camera calibration, usees many assumptions that limit its appicability. Moreover, the implication of using such a model leads to techniques which suffer from the assumption and the implied problem structure induced by the model. The reason behind the inadequacy of the traditional mathematical model for camera representation is not only in per-formance side such as resolution, speed and accuracy but also more systemic in nature e.g. the trend to use a highly distorted fish-eye lens camera to get wider view angle and the use of non-standard planar surface for interaction to create different user experience. In this thesis, we analyze the above challenge in order to facilitate the design, construction and implementation of efficient use of computer vision technology in general. In particular, the thesis begins with more focus to the domain of tabletop augmented reality application. It presents the idea to use universal approximation feature of neural network to solve visual registration problem on virtual interaction surface. Additionally, other kind of visual registration problems also are ad-dressed using the extension of the basic idea. More general usage of the proposed framework in more general computer vision especially in stereo vision is illustrated. The experiment shows that the proposed framework can accommodate a wider class of problems, with more simple and clear setup to the traditional approach.