UC Riverside

This dissertation presents a suite of novel methodologies designed to advance multi-user augmented reality (AR) systems by addressing challenges in localization, mapping, and real-time collaboration. Key contributions focus on enhancing visual-inertial odometry (VIO) and introducing infrastructure-less cooperative SLAM techniques.

Firstly, a Point-Line Cooperative Visual-Inertial Odometry (PL-CVIO) framework is proposed to improve localization accuracy, particularly in low-feature environments. By integrating point and line features and enabling feature sharing between neighboring robots, PL-CVIO leverages geometric constraints to achieve robust, cooperative localization. The framework employs covariance intersection (CI) to ensure consistent state estimation across multiple agents.

Secondly, a novel map-assisted VIO system is introduced by leveraging Neural Radiance Fields (NeRF) to encode compact and photorealistic 3D maps. These maps provide robust geometric constraints for localization, addressing key challenges such as pose initialization, drift correction, and environmental adaptability. A pose initialization model is proposed by using geodesic errors. Besides, an online VIO algorithm is developed, which leverages both real-world and NeRF-rendered images to update the state, demonstrating significant improvements in accuracy and robustness.

Thirdly, we propose CooperSLAM, a lightweight, infrastructure-free cooperative SLAM algorithm designed for multi-user AR in dynamic and resource-limited environments. CooperSLAM enables efficient peer-to-peer communication and sparse map feature sharing, enhancing scalability while reducing bandwidth requirements. By decoupling map points and key frames and introducing opportunistic relocalization strategies, CooperSLAM facilitates effective collaboration without reliance on centralized infrastructure.

Extensive simulations and real-world experiments validate the performance of the proposed methods. Results demonstrate substantial improvements in localization accuracy, robustness, and scalability compared to existing methods. This work contributes to the development of intelligent, collaborative AR systems designed to function effectively in dynamic and infrastructure-less environments, offering potential applications in immersive technologies, robotics, and related fields.