Efficient online continual learning in sensor-based human activity recognition
Loading...
URL
Journal Title
Journal ISSN
Volume Title
School of Electrical Engineering |
Master's thesis
Unless otherwise stated, all rights belong to the author. You may download, display and print this publication for Your own personal use. Commercial use is prohibited.
Authors
Date
2024-08-31
Department
Major/Subject
Communications Engineering
Mcode
ELEC3029
Degree programme
Master's Programme in Computer, Communication and Information Sciences
Language
en
Pages
66
Series
Abstract
Sensor-based human activity recognition (HAR) has been playing an important role in applications such as personal health monitoring and smart homes. In real-world scenarios, machine learning models for HAR are often needed to operate in dynamic environments, where people may perform the same activities in different patterns or conduct new activities post-deployment. This thesis focuses on Online Continual Learning (OCL), where sensors collect data in streaming format, and models must adjust to the changes that occurred in streaming data such as new classes and new patterns, while retaining old knowledge. In addition, HAR models are often deployed on resource-constrained edge devices such as smartphones and the Raspberry Pi, making resource efficiency a critical necessity. This thesis proposes PTRN-HAR, a resource-efficient solution for OCL that only requires a few labeled new class samples. This model utilizes a pre-trained model to extract features and fine-tunes a Relation Network on streaming data based on the features of replay data and new class data samples. Experiments with three open datasets demonstrate that PTRN-HAR achieves higher performance in OCL scenarios compared to baseline methods, while reducing memory consumption and computational cost.Description
Supervisor
Xiao, YuThesis advisor
Souza Leite, ClaytonKeywords
human activity recognition, deep learning, continual learning, resource efficiency, data efficiency, relation network