Author ORCID Identifier

0000-0003-0487-3709

Defense Date

2026

Document Type

Thesis

Degree Name

Doctor of Philosophy

Department

Computer Science

First Advisor

Dr. Eyuphan Bulut

Second Advisor

Dr. Peter Pidcoe

Third Advisor

Dr. Jane Chung

Fourth Advisor

Dr. Kemal Akkaya

Fifth Advisor

Dr. Tamer Nadeem

Sixth Advisor

Dr. Changqing Luo

Abstract

Wireless sensing has emerged as a transformative paradigm that leverages ubiquitous radio-frequency signals for environmental perception and human activity recognition beyond their primary role in data communication. Unlike wearable sensors and camera-based systems, wireless sensing offers a non-intrusive, privacy-preserving, and cost-effective alternative that operates seamlessly under low-lighting conditions, through walls, and in non-line-of-sight (NLoS) scenarios. The growing availability of edge computing devices further enables real-time on-device data processing, making wireless sensing solutions highly suitable for in-home health monitoring and smart living applications. This dissertation investigates real-life health applications of wireless sensing through four completed systems, each addressing a distinct challenge in the deployment of ambient intelligence for healthcare and activity monitoring. The first contribution is Wi-PT-Hand, a low-cost, end-to-end device-free system xvii for physical rehabilitation tracking of hand and finger movements. Using commercial off-the-shelf WiFi modules and deep learning models optimized for edge deployment, the system performs real-time segmentation, classification, and repetition counting of hand exercises, addressing a critical need in post-stroke and neurological rehabilitation. The system was subsequently extended into a clinically validated, spatially dy- namic platform incorporating a Faraday-enclosed multi-link hardware setup, a speed- invariant signal processing pipeline combining Dynamic Time Warping (DTW) align- ment with drop-frame data augmentation, and a validation-weighted convolutional neural network (CNN) ensemble with temporal majority voting. Experiments with 40 healthy volunteers and 11 patients diagnosed with Parkinson’s disease or stroke yielded classification accuracies of 94.2%, 91.4%, and 86.3% on internal validation, external test, and clinical cohort datasets, respectively, with sub-300 ms inference on a Raspberry Pi 4B. The second contribution extends to daily activity monitoring in senior housing environments. By deploying ESP32-based transceivers and Raspberry Pi modules in real apartment and home environments, the system collected naturalistic long- duration CSI datasets from 11 senior participants over five to nine days each, including participants with Parkinson’s disease diagnoses. The deployment addressed practical challenges including signal interference, environmental changes, and user variability, and used supervised contrastive learning and stacking ensemble models to achieve cross-person and cross-environment generalization. The third contribution introduces MockiFi, a novel data augmentation framework using Conditional Neural Processes (CNPs) for zero-shot learning. The approach generates synthetic Channel State Information (CSI) data for unseen user-activity pairs from a single reference activity of a new individual, using activity-to-activity transformation patterns learned from known users. Classifiers trained exclusively on MockiFi-generated data achieved accuracies within one percentage point of classifiers trained on real data, significantly reducing the burden of data collection and labeling in new deployments. The fourth contribution, AgnoGest, addresses a fundamental interference bottleneck in real-world wireless gesture recognition, i.e., the problem of distinguishing subtle hand gestures from dominant concurrent full-body locomotion. AgnoGest is an edge-optimized, locomotion-agnostic multimodal sensing platform that fuses 60 GHz mmWave radar point-cloud data and dual-link WiFi CSI on a Raspberry Pi 4B. A triple-adversarial learning architecture using Gradient Reversal Layers (GRL) suppresses locomotion-specific signatures across all three sensor streams, while dual- attention gated fusion dynamically weighs each modality’s contribution to gesture and locomotion classification. The system achieves an average gesture recognition accuracy of 87.64% across entirely unseen volunteers over nine locomotion conditions, demonstrating that multimodal fusion of mmWave and WiFi substantially outperforms either modality alone in dynamic, real-world settings. Together, these efforts establish a comprehensive path toward scalable, privacy- aware, and user-adaptive wireless sensing solutions for healthcare, assisted living, and personalized activity monitoring, with particular emphasis on edge deployability, clinical validity, and multimodal robustness.

Rights

© The Author

Is Part Of

VCU University Archives

Is Part Of

VCU Theses and Dissertations

Date of Submission

8-10-2026

Share

COinS