| Abstract: | Augmented reality is a young yet a very active research area where the registration problem is a key issue. Registration in augmented reality is the process of aligning a virtual object with a real scene in the three-dimensional space. Solving the registration problem will yield more realistic and practical systems for augmented reality. Computer vision techniques remain the most promising way for solving the registration problem. In this paper, an augmented reality system that makes use of recent advancements in computer vision research is presented. The system is practical and flexible to be used in real-life situations. In particular, the user intervention has been kept minimal, intuitive and does not require any special device nor any knowledge in computer vision. The new registration method we present uses a target plane defined by three points in the scene. In particular, the plane does not need to be known and can be an actual surface or a virtual one. In addition to the registration of the virtual object in the first pair of images, our method uses trifocal tensor to automatically propagate the registration to the remaining images of the sequence. The proposed method is demonstrated by our experiments that have the feasibility and robustness of the three stages of our system: calibration, registration and tracking.
|