FSTNet: Learning spatial-temporal correlations from fingerprints for indoor positioning

Due to the absence of a global positioning system (GPS) signal and pervasive deployment of wireless local area networks (WLANs), Wireless Fidelity (WiFi) fingerprinting has been attracting much attention in indoor positioning services. However, accurate indoor positioning is challenging due to fingerprint ambiguity and instability problems. Most existing positioning methods do not model the spatial and temporal correlations of fingerprints, and this cannot yield satisfactory localization results. Targeting the shortcomings of existing studies, in this paper, we propose a novel deep learning framework, named FSTNet, to learn the spatial- temporal correlations from fingerprints to improve indoor positioning accuracy. In our framework, we first propose a new concept called path fingerprint to solve the fingerprint ambiguity and instability problem. Then, a convolution network is utilized to efficiently capture the local features in path fingerprints. Next, the fingerprint attention mechanism is designed to efficiently capture the spatial features and obtain stable positioning results. Finally, actual on-site experiments are conducted to verify the effectiveness of FSTNet. It is concluded that the proposed modeling and positioning method can effectively capture the temporal and spatial correlations in the received signal strength (RSS) measurements to improve positioning accuracy. In particular, the proposed model could achieve a performance improvement of 44% in terms of a mean positioning error, and 99.2% of the positioning errors are within 2 m.

Automated summary

Due to the absence of GPS signal and the popularity of WLANs, WiFi fingerprinting has attracted attention in indoor positioning. However, accurate indoor positioning is challenging due to fingerprint ambiguity and instability. This paper proposes a deep learning framework called FSTNet to learn the spatial-temporal correlations from fingerprints and improve indoor positioning accuracy. Experimental results show that FSTNet effectively captures the temporal and spatial correlations in RSS measurements, achieving a 44% improvement in mean positioning error and 99.2% of positioning errors within 2 m.