Cost Reduction in Fingerprint-Based Indoor Localization using Generative Adversarial Network

Abstract

Among the various methods of Wi-Fi localization, fingerprint-based indoor localization is a method with relatively high accuracy. However, it takes a lot of time and labor to initially build an offline database. In particular, as the size of the space increases, the number of reference points to be investigated also increases, so the cost increases proportionally. To solve this problem, a few recent studies have used machine learning to reduce the number of samples collected at each reference point. In this paper, we propose a method to reduce the number of reference points that need to be investigated. We collect channel state information (CSI) at a few reference points sparsely and convert it into a CSI amplitude image. Then, the generative adversarial network (GAN) interpolates the values between each reference point by learning these CSI amplitude images and walking in the latent space to synthesize a dense fingerprint map for improved localization accuracy. Our preliminary results show that GAN can create artificial CSI amplitude data that are similar to the actually measured samples for locations in between trained reference points. © 2021 IEEE.