모델 기반 카메라 추적에서 3차원 모델링의 허용 오차 범위에 대한 분석

Abstract

Camera tracking is a process that estimates camera poses, positions and orientations, relative to target objects or scenes. Accurate and robust camera pose is essential for seamless registration of virtual objects in real spaces. In this thesis, a number of significant works related to augmented reality and camera tracking, including a brief introduction of modeling, are covered as a preliminary. In particular, there are two main issues focused on in this thesis: tolerance analysis of camera tracking in terms of modeling errors; one is through a mathematical model and the other is through a subjective evaluation. In general, model-based camera tracking is robust and flexible against different environment because it uses 3-D knowledge of target objects or scenes for estimating camera poses, which is reconstructed as 3-D models using optical equipment or vision-based methods. Hence, accuracy of 3-D models is important particularly in model-based camera tracking as it directly affects tracking reliability. There have been several methods for the accurate modeling; however, most precise modeling requires procedures. In spite of their efficiency, the accuracy of 3-D models should be carefully considered for the performance of the camera tracking since the error between the created model and camera tracking is still sensitive. In the thesis, a tracking accuracy in terms of modeling errors is explored in two aspects. First, the tolerance of modeling errors is analyzed through a mathematical model, mainly based on Cramer-Rao lower bound (CRLB). Using Cramer-Rao lower bound, the tendency of the variance of computational errors and its Cramer-Rao lower bound are discussed. Second, a subjective evaluation is performed through a user evaluation. Using user evaluation, computational errors of model-based camera tracking and its human perceptual errors are compared, and a permissible range of modeling errors for reasonable tracking is discussed. To summarize, the thesis mainly discusses the efficiency of model-based camera tracking and introduces two approaches for a tolerance analysis of modeling errors: one is based on Cramer-Rao lower bound, the other is based on user evaluation.