증강현실을 위한 모델 기반 카메라 추적

Abstract

Augmented Reality (AR) is computer vision and graphics technology that provides users with intuitive information seamlessly and effectively by overlapping 3-dimensional virtual objects in real space. This technology enables humans to generate additional virtual spaces in real space and to feel reality in mixed virtual and real environment. Thus, space and time limitations can be removed, and the range of human activities is expanded as well. Recently, AR is becoming a sensation because of its capabilities on mobile devices (e.g. smart phone). To generate AR environment, reliable camera tracking for robust registration of virtual objects and effective authoring or modeling method need to be guaranteed. These techniques are active research issues in the field of AR. Camera tracking is the process of camera pose estimation that obtains the translation and orientation matrix of the camera using environment information; thus, accuracy of camera pose is very important for seamless registration of virtual objects to real spaces. This thesis explores a number of significant works related with AR and several camera pose estimation methods as a preliminary. In particular, there are three main issues focused on in this thesis: camera tracking method using two visual cues (features and edges) for AR, interactive modeling for model-based (edge-based) camera tracking, and analysis of tracking accuracy in terms of model error. These issues are written in detail in technical point of view. In general, feature- and edge-based camera tracking is more robust and flexible against different environments because each visual cue is complementary to estimate camera pose. For edge-based camera tracking, pre-defined CAD model is needed, and this model can be created efficiently by interactive modeling. In other words, interactive modeling enhances the efficiency of the camera tracking using two visual cues due to its capability of intuitive 3D target object modeling for edge-based tracking. Therefore, camera tracking using two visual cues is reinforced with interactive modeling in the aspect of usability. In spite of its efficiency, there is error sensitivity between the created model and camera tracking. In the thesis, the variation of the tracking error is calculated and analyzed. A general tendency of the tracking error according to the model error is provided through the analysis of the errors using Cramer-Rao lower bound. To summarize, the thesis discusses approaches to implement AR environment more effectively, e.g. camera tracking using visual cues with interactive modeling, and analyzes the efficiency by visualizing the sensitivity of the errors.