Adaptive Device-Free Localization in Dynamic Environments Through Adaptive Neural Networks

Abstract

Device-free localization based on WiFi Channel State Information (CSI) has attracted considerable attention in recent years. Many of the schemes are learning-based, among which CSI fingerprinting based on deep learning is the most promising. Learning-based schemes assume the distributions of CSI fingerprints keep relatively stable over time. However, this assumption often does not hold. As WiFi signals are prone to be influenced by various environmental factors, CSI fingerprints vary significantly with environmental changes. Consequently, the localization model built for the original environment degrades dramatically in the changed environment. Recalibrating the whole area of interest is labor-intensive and time-consuming. This is a major challenge of device-free WiFi localization. To mitigate this issue, we propose the method AdapLoc, aiming to adapt the original localization model to the changed environment with significantly reduced recalibration effort. AdapLoc is based on one-dimensional Convolutional Neural Network (1D-CNN) and exploits Domain Adaptation (DA) with Semantic Alignment (SA) to achieve adaptation. Extensive evaluations in four real-world single-link testbeds with multiple environmental changes demonstrate the effectiveness of AdapLoc in coping with localization in dynamic environments, outperforming the existing work with respect to localization and adaptation.