Advancing Augmented Reality: Localization Algorithm Analysis and Pre-Positioning Virtual Objects

Abstract

This thesis evaluates localization algorithms for Augmented Reality (AR) applications, focusing on five state-of-the-art monocular-inertial localization algorithms—OpenVINS, VINS-Mono, ORB-SLAM3, Kimera-VIO, and DM-VIO. These algorithms were assessed using publicly available datasets (EuRoC) and custom datasets collected with handheld devices, simulating typical AR user movements. The evaluation highlights trade-offs in accuracy, robustness, and initialization time, providing insights into their suitability for various AR scenarios. A comparative analysis with Google’s ARCore reveals that while custom algorithms have higher precision in outdoor environments, ARCore demonstrates superior precision indoors. A significant contribution of this work is the development of an AR pipeline capable of accurately rendering virtual assets in their intended real-world locations without relying on pre-existing 3D maps. The pipeline comprises four threads: data capture, origin setting, localization, and rendering. It incorporates fiducial markers such as AprilTags to seamlessly align the real and virtual worlds by establishing a shared origin between them.