Prior-Guided Self-Supervised Learning for WiFi-based Human Activity Recognition

Abstract

The advancement of WiFi devices has made it feasible to repurpose WiFi signals for integrated sensing and communication purposes. The proliferation of WiFi devices in our daily lives generates data at an unprecedented scale, providing valuable insights into human behavior and activities. This collected data can be utilized for human activity recognition (HAR) in various applications, particularly in healthcare, where challenges posed by an aging society are increasingly pronounced. Learning-based methods, such as deep learning approaches, have garnered increasing attention and have demonstrated outstanding performance. However,they are inherently data-hungry, presenting challenges due to the privacy and complexity associated with WiFi channel state information (CSI) data.Concurrently, Self-Supervised Learning (SSL) has emerged as a promising approach to tackle the challenge of insufficient labeled data, as observed in other fields such as computer vision. However, its application to CSI-based human sensing tasks remains under-explored.The inherent modality differences between CSI data and other formats, such as vision data,often lead to sub optimal performance when directly applying SSL algorithms developed for these formats to CSI data. This is primarily because these algorithms are designed with the specific characteristics and prior knowledge of their original data format, lacking the tailored insights required for effective CSI data processing. Therefore, an effective approach is to incorporate domain-specific knowledge of CSI data into SSL algorithms as priors to enhance their performance.In this thesis titled "Prior-Guided Self-Supervised Learning for WiFi-based Human Activity Recognition," we conduct a comprehensive and systematic study of existing SSLmethods’ performance on CSI-based HAR tasks. Additionally, we introduce novel SSL techniques with prior knowledge of CSI data that achieve state-of-the-art performance in both in-domain and cross-domain settings. In this context, "in-domain" refers to scenarios where the training and testing data are collected under the same conditions. Conversely,"cross-domain" denotes situations where the training data and testing data are gathered under different conditions. To extract both temporal and spatial features from CSI data, we propose a novel algorithm that incorporates multiple subcarrier characteristics into contrastive learning. DualConFi demonstrates a slight improvement over supervised approaches across three datasets in in-domain scenarios. Moreover, in cross-domain scenarios, DualConFi out performs both supervised and unsupervised baselines. Furthermore, we introduce a neasily implementable plug-in SSL module called ARC that leverages the consistency among antennas of the same device to extract disentangled features of the dynamic path from CSI data. ARC exhibits versatile applicability across various types of SSL algorithms for CSI data, designed to preserve information from the input space while introducing robustness to real-world noise. The integration of ARC into these algorithms enhances their capacity extracting meaningful features while mitigating the impact of environmental uncertainties.This theis validate the effectiveness of ARC through a comprehensive set of experiments,showcasing its capability to enhance algorithm performance in WiFi-based HAR in both in-domain and cross-domain settings. In the in-domain setting, a maximum accuracy of 94.87% (0.5% higher than the supervised baseline) is achieved.

Date of Award	Sept 2024
Original language	English
Awarding Institution	Coventry University
Supervisor	Jiangtao Wang (Supervisor), Hongyuan Zhu (Supervisor) & Dingchang Zheng (Supervisor)