GAN Based Data Augmentation for Indoor Localization Using Labeled and Unlabeled Data

Abstract

Machine learning techniques allow accurate indoor localization with low online complexity. However, a large amount of collected data samples is needed to properly train a deep neural network (DNN) model used for localization. In this paper, we propose to generate fake fingerprints using generative adversarial networks (GANs) based on a small amount of collected data samples. We consider an indoor scenario where collected labeled data samples are rare and insufficient to generate fake samples of a good multitude and diversity in order to provide a good localization accuracy. Thus, both labeled and unlabeled fingerprints are provided to the GAN so that more realistic fake data samples are generated. Then, a DNN model is trained on mixed dataset comprising real collected labeled and pseudo-labeled fingerprints as well as fake generated pseudo-labeled fingerprints. The data augmentation based on real measurements leads to a mean localization accuracy improvement of 9.66% in comparison to the conventional semi-supervised localization algorithm.