Joint Task-Offloading and Communication Framework for Wireless Metaverse
Jankovic, Branislava
Jankovic, Branislava
Author
Supervisor
Department
Computer Vision
Embargo End Date
2025-05-30
Type
Thesis
Date
2025
License
Language
English
Collections
Research Projects
Organizational Units
Journal Issue
Abstract
This thesis addresses the challenges of disaster detection and response by integrating metaverse technologies, Unmanned Aerial Vehicle (UAV) networks, and advanced deep learning techniques into a unified, three-part framework. The first part introduces a novel task-offloading framework tailored for metaverse empowered wireless systems. Using iterative optimization, the framework minimizes a cost function that balances transmission energy and latency. Specifically, convex optimization is applied to time division multiple access for efficient time allocation, while a one-to-many matching strategy ensures effective task offloading. Additionally, the metaverse sensing interval is optimized to improve responsiveness. Simulation results confirm the superiority of the proposed approach over traditional schemes, demonstrating improved cost efficiency across different configurations and scaling conditions. The second part presents a UAV-powered edge computing system for disaster detection, built around a transformer-based deep learning model fine-tuned for real-time aerial image classification. The model is optimized through post training quantization and made interpretable with Explainable AI (XAI) techniques, which visually highlight the key regions in input frames used during prediction. To overcome the limitations of existing disaster datasets, a new dataset - DisasterEye - was developed, featuring diverse geolocations and UAV-captured disaster scenes. The proposed solution reduces inference time and memory usage without compromising accuracy, making it viable for deployment on resource-constrained UAV platforms. The third part introduces a reasoning-based video disaster detection framework, leveraging the Video Analysis Server. This framework includes two phases. In the learning phase, we generate the descriptions for video frames, and based on these descriptions and the known video labels, we generate rules for the classification of future unlabeled videos. In the inference phase, the rules guide the LLM in making accurate predictions. This framework not only produces accurate predictions of the disaster type but also supports them with reasoning. A novel video disaster dataset has been developed to support the framework, featuring nine complex disaster scenarios paired with descriptive text. Extensive experiments show that this method significantly outperforms conventional video classification models, achieving improvements of 3.22%, 0.26%, 3.22%, and 3.07% in accuracy, precision, recall, and F1-score, respectively.
Citation
Branislava Jankovic, “Joint Task-Offloading and Communication Framework for Wireless Metaverse,” Master of Science thesis, Computer Vision, MBZUAI, 2025.
Source
Conference
Keywords
Disaster Detection, Metaverse, Transformers, Large Language Model (LLM), VLM