WiFi-Based Person-Centric Sensing

Abstract

WiFi-based Person-Centric Sensing (PCS) offers a visual privacy-preserving alternative to optical methods by enabling passive, contactless monitoring through existing wireless infrastructure. Its low cost, unobtrusive nature, and wall-penetrating capability make it well suited for large-scale indoor monitoring applications. However, practical deployment remains constrained by data scarcity, limited sensing range of consumer off-the-shelf (COTS) hardware, computational inefficiencies, poor cross-domain generalization, and the abstract, non-intuitive nature of WiFi signals. To address the data scarcity, five publicly available WiFi Channel State Information (CSI)-based PCS datasets are contributed: TOA, Wallhack1.8k, HALOC, 3DO, and WiFiCam. Each dataset targets distinct challenges in long-range and through-wall sensing, domain generalization, and crossmodal translation. Building on this foundation, it is demonstrated that directional sensing with low-cost COTS WiFi systems enables long-range through-wall PCS. Experiments confirm robust presence detection, activity recognition, and localization up to 20 meters and across multiple rooms with a single-link setup, validating the effectiveness of the proposed directional sensing approach in complex indoor environments. To support real-time inference under resource constraints, WiFlexFormer, a lightweight Transformer architecture tailored to the temporal and spectral characteristics of WiFi CSI, is introduced. With only ≈ 50k parameters, it achieves inference latencies of ≈ 10 ms on embedded hardware while matching or surpassing the performance of significantly larger generic vision and RF-specific architectures. To improve robustness across domains, data augmentation and preprocessing strategies that enhance generalization without target-domain access are investigated. Building on these insights, the Domain-Adversarial Test-Time Adaptation (DATTA) framework is proposed. DATTA leverages domain-adversarial training, test-time adaptation, random weight resetting, and data augmentation to enable robust, real-time adaptation to domain shifts, achieving state-of-the-art cross-domain generalization performance. Lastly, the thesis presents the first approach to synthesize RGB images from WiFi CSI in through-wall scenarios. The WiFiCam architecture, based on a multimodal variational autoencoder, reconstructs coherent, semantically meaningful images, enabling camera-free visual monitoring and improving the interpretability of CSI for downstream tasks. These contributions address core limitations in data, hardware, efficiency, generalization, and interpretability, advancing WiFi-based PCS toward scalable real-world deployment.