Citation

Salehi, Parsa (2026) Adaptive novelty detection system for continual learning of human activity. Masters thesis, Multimedia University.

Full text not available from this repository.

Abstract

Human activity recognition in smart homes has attracted an increasing attention from the research community due to the important role of smart homes in supporting its users, typically the elderly who are living alone. However, current activity recognition systems are impractical for real-world applications since they assume a closed-set setting. Human activities are dynamic in nature and can change over time. Existing incremental learning methods for novelty detection have downsides of frequent model retraining or high memory overhead. This thesis proposes an adaptive activity recognition system based on prediction by partial matching that enables incremental learning of new activities without full model retraining. The proposed approach dynamically updates the base model by incorporating new activity while retaining previously learned ones. Experiments were conducted on publicly available smart home datasets, which demonstrates the system’s ability to continuously learn through incremental updates, avoiding cost and disruption of rebuilding model from scratch. Our method achieved comparable results compared to batch learning. Although many activity recognition systems based on supervised learning have been proposed over the years, one of the problems with supervised learning is that it requires a large amount of labelled data for training. In the context of activity recognition in smart homes, most of the labelling tasks are often done manually by the user themselves. Manual data labelling is not only tedious but also time-consuming and error prone. There are studies that attempt to use active learning methods for data annotation. However, these methods still require some amount of effort from the user and labelled data. In this thesis, we propose an automated labelling approach based on transformer-based deep learning method, which leverages on spatio-temporal information. Experimental results showed that our method can label the majority of the activities without relying on prior labelled data.

Item Type:	Thesis (Masters)
Additional Information:	Call No.: TK7882.P3 S25 2026
Uncontrolled Keywords:	Human activity recognition
Subjects:	T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK7800-8360 Electronics
Divisions:	Faculty of Computing and Informatics (FCI)
Depositing User:	Ms Nurul Iqtiani Ahmad
Date Deposited:	05 Oct 2026 03:31
Last Modified:	05 Oct 2026 03:31
URII:	http://shdl.mmu.edu.my/id/eprint/16868

Downloads

Edit (login required)