More Accuracy Less Fingerprints: Wi-Fi Indoor Localization via Generative Adversarial Networks

High-cost site survey is one of the bottlenecks of Wi-Fi fingerprinting indoor localization. Rather than leverage inertial sensors or floorplan, we propose LocGAN to generate virtual fingerprints (VFPs) with only a small number of labeled (with locations) and yet a large number of unlabeled ones. To this end, LocGAN is a semi-supervised deep generative model consisting of TriReg, encoder, generator, and discriminator. As a tri-net based regressor, TriReg provides pseudo-labels for unlabeled fingerprints. Under a generative adversarial network (GAN) framework, LocGAN is able to learn underlying distributions of fingerprints from both labeled and unlabeled ones thus generating high-accuracy VFPs. We also design several effective training strategies to further improve its performance. To evaluate LocGAN, we prototype a Wi-Fi indoor localization system based on it. Extensive experiments are carried out in real-world scenarios with areas over 8,200m ² . The experiment results demonstrate that compared with the state-of-the-art counterparts, LocGAN achieves more accuracy with less labeled fingerprints, reducing the cost of site survey significantly.