Loading...
Quality enhancement for mobile augmented reality
Muthu Kumara Swamy, Shneka
Muthu Kumara Swamy, Shneka
Citations
Altmetric:
Advisor
Editor
Date
Date Issued
2024
Date Submitted
Collections
Research Projects
Organizational Units
Journal Issue
Embargo Expires
2026-04-04
Abstract
Mobile Augmented Reality (MAR) enhances user AR experiences on smartphones through two main forms of overlay: segmented image overlay and virtual object overlay. The quality of these overlays is crucial for a seamless MAR experience. This dissertation explores methods to measure and enhance the quality of MAR to improve user experience, focusing on developing real-time, lightweight techniques suitable for the computational and memory constrained smartphones. To support better segmented image overlay, we design and develop SegQNet, a real-time method employing a lightweight Convolutional Neural Network to assess segmentation quality without relying on ground truth. For virtual object placement, we reduce spatial drift by addressing memory constraints of smartphones and real-time requirements of MAR applications using two approaches. First, we design an online method using depth images to measure spatial drift of virtual objects and then use that as feedback to reduce the drift by providing more accurate pose estimation. This is achieved by using the RGB-D SLAM method (more accurate but also more costly) when the virtual object is visible and switching to the Monocular SLAM method (less accurate but also less costly) when the virtual object is not visible. Second, we demonstrate that dynamically switching between computationally intensive Simultaneous Localization and Mapping (SLAM) and less resource-intensive Visual Odometry (VO) algorithms can significantly reduce memory usage and latency while maintaining pose estimation accuracy. Our findings also underscore the importance of metrics like accuracy, precision, and recall in detecting virtual objects as critical quality indicators for MAR. While these methods are effective in stationary environments, they may not perform well in dynamic scenes. To address this, we develop a real-time pose estimation algorithm for dynamic objects in MAR, integrating segmentation, homographic transformation, and template matching. Furthermore, recognizing the limitations of single-user scenarios, we develop a multi-user MAR system that improves spatial consistency of virtual objects via more accurate pose estimation. This is achieved by selecting the right keyframes for coordinate transformation among users and use energy-aware collaboration among users. Overall, this dissertation presents a comprehensive approach to improving MAR quality, balancing computational efficiency with high performance on resource-constrained mobile devices.
Associated Publications
Rights
Copyright of the original work is retained by the author.