University of Twente Student Theses


Federated Learning for Indoor Human Activity Recognition: Adapting to Changing Realistic Environments

Linden, M.M. van der (2023) Federated Learning for Indoor Human Activity Recognition: Adapting to Changing Realistic Environments.

[img] PDF
Abstract:This work focuses on the optimisation of federated learning (FL) for indoor human activity recognition (HAR) in dynamic environments using Wi-Fi-based channel state information (CSI) signals and resource-constrained devices. This is achieved by analysing the impact of various parameters to deal with unseen activity locations and dynamic client participation. Promising advancements are found in the amount and type of communication and local computation of clients, as well as the use of existing models to increase convergence speed of models based on unseen data. Traditional HAR methods can be intrusive, are limited by environmental constraints and invade ones privacy. FL, a distributed learning technique, allows for collaboration between multiple clients with each their own unique data, which maintains privacy while allowing for a powerful machine learning (ML) model. The main limitation of FL for HAR found in previous work is the lack of realistic testing environments. By simulating scenarios with resource-constrained devices, this study explores the impact of local computation, varying aggregation algorithms, limited data availability and changing environments. The research highlights opportunities and challenges of realistic environments based on extensive analysis and comparison of the varying parameters, providing recommendations to maximise the benefits from using FL to train a deep learning model for indoor HAR. The thesis concludes with an outlook for the future direction of this system and an overview of the remaining challenges to be addressed.
Item Type:Essay (Master)
Faculty:EEMCS: Electrical Engineering, Mathematics and Computer Science
Subject:54 computer science
Programme:Embedded Systems MSc (60331)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page