|M.Sc Student||Keren Sara|
|Subject||An Augmented Reality System for Uncalibrated Images|
of Indoor Architectural Scenes
|Department||Department of Electrical Engineering||Supervisors||Professor Ayellet Tal|
|Professor Ilan Shimshoni|
An augmented reality system is an environment in which virtual objects are overlaid upon the user’s view of the real world. It should provide a convincing perception of the combination of the virtual world and the real world. One of the major problems an augmented reality system should solve is camera calibration and virtual object registration. Our work discusses the problem of inserting three-dimensional models into a single image of an indoor architectural scene. The image is taken by a camera whose internal calibration parameters and whose position and orientation in the scene are unknown. Typically, in architectural scenes lines appearing in the images are parallel to the three major axes. We utilize this information from a single image to recover the parameters of the real camera, which are used in overlaying the virtual objects in their correct positions and orientations in the real image. The real image as given by the user is preserved. An important aspect of our work is a theoretical and an experimental analysis of the errors of the camera’s parameters and of the errors of the projection of the virtual objects. In our augmented reality system the user chooses a location in the image for the object to be placed in. The system implements the placement of the virtual objects upon the image without constructing a full 3D model of the real scene. Therefore, it needs minimal information to be supplied by the user to perform the task. Our goal is to determine whether the results are accurate enough so that a human viewer will not be able to detect differences between the size, the position and the orientation of a real object and of its 3D model inserted into the scene. Our analysis and experiments have shown that errors in the placement of the objects are un-noticeable by the viewers. We also reached some conclusions on the recommended camera orientation and location in the scene that would give the best results in the recovering of the camera parameters and in the overlaying of the virtual objects upon the real image.