Document Type

Thesis

Disciplines

1.2 COMPUTER AND INFORMATION SCIENCE, Computer Sciences

Abstract

WiFi-based Human Activity Recognition (HAR) has emerged as a promising, device-free, and privacy-preserving alternative to vision- and wearable-based systems. Its ability to function without cameras or body-worn sensors makes it highly suitable for applications in healthcare (e.g., fall detection, patient vital-sign monitoring, and elderly care), smart homes, and security surveillance. Despite these advantages, WiFi-based HAR systems often struggle to generalize across different users and environments due to domain shifts and subject bias, limiting their real-world applicability. The aim of this thesis is to develop adaptive, device-free, and privacy-preserving WiFi-based HAR methods that improve cross-user and cross-environment performance, while systematically analyzing the effects of subject variability and spatial diversity on recognition accuracy. In the initial stage of this research, I develop and evaluate a baseline HAR system using Channel State Information (CSI), forming a solid foundation for activity classification. I then conduct cross-domain evaluations to examine how well the system performs under variations in users and environments. To enhance adaptability, I apply transfer learning techniques, which offer modest improvements but also expose significant shortcomings. This leads to a focused analysis of subject bias in WiFi-based HAR, where I demonstrate that conventional evaluation methods often overestimate performance by failing to account for individual variability. As retraining models for each new deployment/domain is costly and inefficient, the second stage focuses on domain adaptation using feature augmentation to address domain variability. Both supervised and semi-supervised augmentation techniques are applied to enrich feature representations and enhance generalization. While effective in improving cross-domain robustness, they rely on labeled data whose collection is often expensive and time-consuming, especially in new or changing environments, an assumption that does not always hold in real-world deployments. To overcome this constraint, the third stage introduces a novel unsupervised domain adaptation approach that integrates the Domain Adversarial Neural Network (DANN) with a feature alignment method. This approach aligns feature distributions between source and target domains without requiring any labeled data from the target, leading to substantial improvements in cross-subject and cross-environment recognition. In the final stage, I collect our own dataset in a real-world living room setting to study the effects of daily changes in an in-situ environment and subject variability. This dataset enables us to analyze how the daily environmental changes and differences in subject properties, such as appearance, influence CSI signal patterns and model performance. Our findings highlight the critical role of spatial diversity and subject variability factors, further validating the need for adaptive and lightweight HAR systems suitable for diverse and dynamic settings. This thesis advances WiFi-based human activity recognition by addressing subject and envi ronmental variability. It establishes a CSI-based baseline system, develops feature-augmentation strategies for domain adaptation, and introduces an unsupervised domain-adaptation method that combines DANN with feature alignment, improving cross-subject and cross-environment perfor mance without labeled target data. Additionally, analysis of a real-world in-situ dataset highlights the impact of daily environmental changes and individual differences, providing actionable insights for designing adaptive, device-free, privacy-preserving HAR systems suitable for dynamic real-world settings.

DOI

https://doi.org/10.21427/3tyj-nb86

Creative Commons License

Creative Commons Attribution-NonCommercial 4.0 International License
This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License

Download


Share

COinS