Advancing Egocentric Action Recognition for Perceptually-enabled Task Guidance
This project aims to create a task guidance assistant using the HoloLens headset that will guide the user through augmented reality. The primary focus of this thesis lies in the integration of an action recognition framework for egocentric videos, crucial for task prediction within the system. Devel...
| Autor: | |
|---|---|
| Tipo de recurso: | tesis de maestría |
| Fecha de publicación: | 2023 |
| País: | España |
| Institución: | Universitat Politècnica de Catalunya (UPC) |
| Repositorio: | UPCommons. Portal del coneixement obert de la UPC |
| Idioma: | inglés |
| OAI Identifier: | oai:upcommons.upc.edu:2117/420290 |
| Acceso en línea: | https://hdl.handle.net/2117/420290 |
| Access Level: | acceso abierto |
| Palabra clave: | Computer vision Augmented reality action recognition computer vision reconeixement d'accions Visió per ordinador Realitat augmentada Àrees temàtiques de la UPC::Informàtica::Intel·ligència artificial |
| Sumario: | This project aims to create a task guidance assistant using the HoloLens headset that will guide the user through augmented reality. The primary focus of this thesis lies in the integration of an action recognition framework for egocentric videos, crucial for task prediction within the system. Development starts in a kitchen environment, with the intention of using transfer learning for military scenarios in the future. Epic-Kitchens serves as an initial reference dataset, subsequently followed by the creation of a customized dataset. Various state-of-the-art action recognition models are considered, with Omnivore being the final choice. Initial results show 14.23% Top 5 action recognition accuracy within the created dataset. Through classifier modifications and application of diverse post-processing video techniques, this accuracy is significantly improved, culminating in an impressive 83.76%. |
|---|