امین رضا غلامی
|
: Abstract
Conventional augmented reality (AR) systems rely on recovering the camera pose at every frame to project virtual object onto the ۲D images. This estimation is generally carried out by Simultaneous Localization and Mapping (SLAM), which assumes that the intrinsic parameters of the camera are fixed or known throughout the sequence. Such restrictions limit the applicability of SLAM-based AR in more dynamic or uncalibrated environments.
The work presented in this dissertation proposes an alternative framework that bypasses explicit camera parameters recovery. Instead of estimating the camera’s parameters, the method exploits epipolar geometry to directly compute two-dimensional projections of virtual objects. The process involves tracking feature points across frames and deriving fundamental matrices between keyframes and newly acquired frames. These matrices then enable the placement of virtual objects by determining the intersections of corresponding epipolar lines.
Because feature correspondences are often contaminated with mismatches, we incorporate strategies to identify and suppress outliers, including a robust scheme that maintains stability even when some of the fundamental matrices are unreliable. To further reduce computational costs, optimizations are introduced that limit the number of RANSAC iterations required. Furthermore, a new z-buffering strategy is introduced, which eliminates the necessity of reconstructing the complete ۳D scene geometry for each frame. The resulting approach is resilient to variations in camera intrinsic parameters over time and demonstrates that reliable AR overlays can be achieved without pose estimation.
|
|
|
|