Dual-Band Wi-Fi Based Indoor Localization via Stacked Denosing Autoencoder

With the ever-increasing demand of location-based services (LBS), Wi-Fi based indoor localization has attracted increasing attentions. This paper is dedicated to addressing two critical problems: a) signal fluctuation due to unforeseeable interferences during the offline training phase; b) insufficient real-time signal measurements at certain point due to the target movement during the online localization phase. Specifically, we first give an intensive analysis on the characteristics of received signal strength indicator (RSSI) in indoor environments with respect to both time-domain and frequency-domain. Then, inspired from the advantages of Stacked Denosing Autoencoder (SDA) in terms of recognizing and stabilizing the original features, we propose a dual-band SDA (DBSDA) based model to create more distinguishable fingerprints by extracting the RSSI features at each reference point (RP). In this model, both 2.4GHz and 5GHz RSSIs are exploited to train the SDA neural network and construct the offline fingerprint database. On this basis, we propose a data generation scheme, which is designed based on the observation that environmental interferences are similar in proximate spots. So, the designed scheme can generate signal values at certain point based on its nearby RSSI measurements when there are not enough inputs for the SDA neural network. Finally, we propose a locally weighted liner regression (LWLR) based method to predict the coordinate of the target. For performance evaluation, we implement the system prototype and give comprehensive experiments in real-world environments, which demonstrate the effectiveness and robustness of the proposed solutions.