Abstract
We present an augmented reality (AR) system that is based on a tabletop system and has a hemisphere omnidirectional camera. We perpendicularly set the camera on a tabletop display and showed an omnidirectional image on the display screen. When users present objects to the camera, these objects are captured and recognized by using a specific object recognition technique. On the bases of the recognition result, AR contents are overlaid onto the omnidirectional image on the screen. Our system includes following two advantages. First, users can interact with each other by recognizing the relationship among the users, objects, and their AR contents because the entire surrounding of the tabletop is shown as a circular image on the screen. Second, multiple users interact here without the need for any specific devices. To describe the applicability of our system, we present various examples of applications exploiting its advantages.
Introduction
Traditional Augmented Reality (AR) systems frequently rely on Head-Mounted Displays (HMDs) or handheld devices. While effective, HMDs can be bulky and isolate users, while handheld AR occupies the user's hands, hindering natural interaction. Tabletop AR systems offer a collaborative alternative but often struggle to capture the full environmental context—specifically the spatial relationships between multiple users and objects—without complex multi-sensor arrays.
This research introduces a tabletop AR system utilizing a single omnidirectional camera to capture a 360-degree view. By addressing the geometric distortion inherent in wide-angle lenses, the system enables seamless, multi-user interaction with digital content without specialized personal hardware.
System Design and Methodology
The hardware configuration consists of a flat-panel LCD display placed horizontally, with an omnidirectional camera (utilizing a hyperbolic mirror) mounted perpendicularly above the center of the screen. This setup allows the display to act as a digital mirror, showing a live circular feed of the table and its surroundings.
To enable robust interaction, the system follows a multi-stage software pipeline:
- Panorama Expansion: The raw, distorted circular image is "unwarped" into a rectangular panorama. This linearization is critical for normalizing the appearance of objects for the vision algorithms.
- Object Recognition: The system employs a Bag-of-Visual-Words (BoVW) approach. SIFT (Scale-Invariant Feature Transform) features are extracted from the panorama and matched against a database using a hierarchical K-Means tree, allowing for the identification of specific items like trading cards.
- Virtual Perspective Projection: To accurately overlay AR graphics, the system maps coordinates from the unwarped panorama back to the circular display view using a virtual perspective projection model.
Experimental Evaluation
The system was evaluated using a dataset of 50 commercial trading cards. Researchers tested recognition stability at various distances and rotation angles. A key technical insight was that performing recognition on the expanded panorama significantly outperformed matching on the raw circular image.
Quantitative results demonstrated that the panorama expansion provides a much higher number of correct correspondence points, leading to more stable object tracking. While the current system achieves a frame rate of approximately 3.37 fps (with a 296ms latency per frame), it successfully maintains recognition even when objects are rotated up to 70 degrees.
Conclusion
This tabletop AR system provides a unique collaborative environment where the entire surrounding area is captured and utilized for interaction. By normalizing omnidirectional distortion through panorama expansion, the system achieves reliable specific object recognition. Future work will focus on reducing latency through GPU acceleration and addressing resolution degradation at the periphery of the hyperbolic mirror's view.
Summary updated