Joint Localization and Radio Map Generation using Transformer Networks with Limited RSS Samples

Abstract

Received Signal Strength (RSS) based fingerprinting method is extensively used for localization in the Internet of Mobile Things (IoMT) environment due to the low cost and ubiquitous nature of Wireless Fidelity (Wi-Fi) signals. However, its wide adoption is limited by the inability of this method to handle temporal RSS variations and mobile device heterogeneity issues. A labor-intensive solution to this problem is collecting large number of RSS fingerprints that capture the variations in the new environment and the new mobile device. In this paper, we propose a RSS fingerprint generation method for the heterogeneous mobile device or new environment by limiting the RSS fingerprint sampling to a few locations. We accomplish this by utilizing a transformer neural network architecture for generating RSS samples. The proposed method with few calibration samples (95 % fewer samples) provides comparable localization performance with the state-of-the-art methods that use the labor-intensive task of RSS sample collection in the dynamic environment for two widely used IoMT testbeds.