Abstract
Camera setup, calibration and visual based registration of Augmented Reality (AR) based tabletop setups can be a really complicated and time-intensive task. Homography is often used liberally despite its assumption for planar surfaces, where the mapping from the camera to the table can be expressed by a simple projective homography. However, this approach often fails in curved and non-planar surface setups. In this paper, we propose a technique that approximates the values and reduces the tracking error-values by the usage of a neural network function. The final result gives a uniform representation of the camera against combinations of camera parameters that will help in the multi-camera setup. We present the advantages with demonstration applications, where a laser pointer spot and a light from the lamp will be tracked in non planar surface. © 2006 Springer-Verlag Berlin/Heidelberg.