Representation learning for human sensing with WiFi channel state information
File(s)
Author(s)
Huang, Shuokang
Type
Thesis
Abstract
Human sensing with radio signals has emerged as a less-intrusive and environment-robust alternative to approaches based on on-body sensors or cameras, and WiFi signals further support device-free detection without the need for dedicated hardware. Specifically, Channel State Information (CSI) monitors WiFi signal variations caused by human interference and accordingly contains useful features for human sensing. Since useful features in CSI are not explicit and typically intertwine with excessive noise, considerable efforts have been devoted to representation learning for implicit human features from CSI. However, several practical challenges still undermine the representation learning with WiFi CSI, such as incomplete CSI samples, multi-user activity sensing, the scarcity of fine-grained labels, and cross-domain human sensing. To tackle incomplete CSI samples, this thesis proposes DiffAR, a temporal-augmented approach to augment CSI samples with diffusion models for advanced WiFi-based human sensing. Considering that most existing methods, including DiffAR, focus on single-user activity sensing due to the lack of multi-user datasets, this thesis constructs the first WiFi-based multi-user activity sensing dataset, WiMANS, posing new challenges while opening up opportunities for future research. According to experiments on WiMANS, the scarcity of fine-grained labels hampers representation learning in multi-user scenarios, and thus this thesis devises CrossAR, a WiFi-Vision approach that adopts unlabeled CSI and auxiliary labels from pre-trained vision models to extract informative features through cross-modal contrastive learning. Despite the above significant progress, cross-domain human sensing with WiFi CSI is still challenging, especially for human pose estimation of which the latest methods are limited to single domains and fall short in cross-subject/environment scenarios, owing to domain-specific confounders. To eliminate these confounders, this thesis presents GenHPE, which formulates generative counterfactuals for domain-independent representation learning with WiFi CSI. Overall, this thesis discusses and handles practical challenges in representation learning with CSI, contributing to more effective and efficient WiFi-based human sensing.
Version
Open Access
Date Issued
2025-08-06
Date Awarded
01/01/2026
License URL
Advisor
McCann, Julie
Publisher Department
Department of Computing
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)