Toward Long-Term Effective and Robust Device-Free Indoor Localization via Channel State Information

Abstract

With the rapid development of location-based IoT applications in recent years, indoor device-free passive localization based on Wi-Fi channel state information (CSI) has attracted considerable attention. In this article, we propose a long-term effective, robust, and accurate device-free passive fingerprinting localization scheme LTLoc, which only requires a single communication link. It takes the amplitudes extracted from the CSI along with the calibrated phases as fingerprints and trains a deep neural network (DNN) regression model to estimate the target location. Since Wi-Fi signals are susceptible to various environmental factors, CSI fingerprints also change over time, making the performance of the localization model built with the fingerprint database drop dramatically over a long period, and recalibrating the entire positioning area is laborious and time-consuming. To address this problem, we design an adaptive DNN (AdaptDNN) based on meta-networks by combining deep learning and domain adaptive methods. It can use meta-network learning to determine which layers and features of the DNN need to be transferred to automatically adapt to CSI fingerprints change. Extensive evaluations in an indoor environment with significantly different CSI fingerprints over six days have shown that LTLoc’s effectiveness in coping with changing CSI fingerprints over a long period is significantly superior to existing work in terms of localization and adaptability.