CSI-Based WiFi Sensing for Respiration Monitoring and Passive Tracking

Abstract

This thesis develops two lightweight and robust WiFi CSI-based frameworks for respiration monitoring and passive localization. For respiration, a dual-branch CNN–Transformer network with attention and contrastive learning discriminates subtle respiratory signals from motion artifacts, achieving over 90% accuracy under mobility and outperforming conventional filtering or frequency-domain methods. For localization, a real-time solution using one transmitter and two receivers employs a CSI quotient model, Doppler features, and differenced AoA/ToF measurements. A geometry-constrained EKF reconstructs sub-meter trajectories, validated across LoS and NLoS scenarios with strong robustness and scalability. Overall, the synergy of deep learning and physical modeling significantly enhances WiFi sensing performance under real-world constraints, demonstrating generalizability on commodity hardware and laying a foundation for smart healthcare and ambient intelligence.