Robust Indoor Localization across Smartphone Models with Ellipsoid Features from Multiple RSSIs

Localization for mobile devices has become important as the basis technology for various ubiquitous computing applications. While GPS is leveraged as the de-facto standard technology in outdoor localization, its accuracy is poor indoors. For twenty years, researchers have tried to investigate indoor localization technology using fingerprinting from received signal strength indicators (RSSIs). With the widespread use of smartphones in the last decade, device dependency (e.g. antenna characteristics) must be considered to avoid performance degradation, while most of the recent localization approaches assume that all the smartphone models have the same device characteristics.

In this paper, we propose a novel feature representation based on multiple RSSIs for compensating performance degradation against smartphone models changes. In contrast to the previous feature representation based on a single RSSI, our new feature representation, which we call Ellipsoid features, employs tuples of pair of RSSIs to eliminate device dependence in the path loss model for wave propagation. In contrast to recent advances in machine learning methods such as domain adaptation, multi-task learning, and semi-supervised learning, our approach requires no additional dataset nor retraining for the new target models. This simplicity would promote ubiquity of indoor localization in the era of smartphones. Moreover, our feature representation works well compared to the state-of-the-arts in feature representations based on multiple RSSIs even when only a small number of access points (APs) are available. Experimental result using smartphone devices including Android Nexus5, Nexus5X, Nexus6P, and Xperia X Performance shows that our approach achieves superior performance over the state-of-the-art indoor localization models as well as robust performance against device changes.