Time matters: Empirical insights into the limits and challenges of temporal generalization in CSI-based Wi-Fi sensing

Abstract

Wi-Fi is ubiquitous, and Channel State Information (CSI)-based sensing has often emerged as superior for tasks like human activity recognition (HAR) and indoor positioning (IP) The foundational premise is that similar scenarios exhibit similar CSI patterns. However, establishing such similarities is challenging due to signal attenuation and multipath effects caused by static and dynamic objects, that create complex interaction phenomena. Although acknowledged in literature, a comprehensive study of how these variables affect CSI patterns across scenarios, particularly their long-term impact on real-world applications, is still missing. In fact, many recent works focus on laboratory settings disregarding temporal generalization when testing their solutions. Here, we present a systematic study of the reliability of CSI-based sensing, consolidating key challenges and insights previously scattered in the literature. We provide a clear and independent perspective about the need of considering temporal aspects when developing CSI-based sensing approaches, particularly for real-world applications. To achieve that, we consider two tasks, IP and HAR, combining theoretical modeling with experiments using state-of-the-art methods. We show how tasks dependent on reflections from static objects, like IP, are severely impacted by disturbances that accumulate over time , also in the absence of physical modifications of the environment. In contrast, those relying on reflections from dynamic objects, like HAR, face fewer challenges. Our findings, supported by novel real-world datasets for CSI fingerprint-based IP and CSI stability analysis over time, suggest that future research must consider time as a crucial factor both in the development and test of approaches.