Leveraging online learning for domain-adaptation in Wi-Fi-based device-free localization

Wi-Fi-based device-free localization (DFL) will be an integral part of many emerging applications, such as smart healthcare and smart homes. One popular approach to DFL in Wi-Fi makes use of fingerprinting based on channel state information (CSI). Unfortunately, high-quality fingerprints cannot easily be obtained in many real-world environments due to the complicated, time-varying and multipath conditions which exist. Additionally, existing methods struggle to update the DFL models in a real-time manner to track changes of environment. To address these issues, an online data-driven modelling DFL framework is designed for robustness enhancement. Specifically, the raw CSI data is first augmented with the hidden layer parameters of an online deep neural network to strengthen the pair-to-pair mappings between signal variations and a target's location. The radio map created with the augmented fingerprints can be updated with new sequential data collected from other domains, such as different times and layouts of the same environment. Subsequently, a novel online DFL model is established using these augmented fingerprints, which itself can be updated with new sequential data from other domains without the need for retraining. A forgetting mechanism is considered to mitigate the effects of outdated data on the localization performance. To validate our new framework, a comprehensive set of experiments have been performed in various environments for different scenarios. The experimental results verify the robustness and responsive tracking ability of the proposed online data-driven modelling DFL framework.