Enhancing Indoor Localization With Semisupervised Teacher-Student Network Using Wi-Fi CSI

We investigate indoor fingerprinting using Wi-Fi channel state information, focusing on two semisupervised strategies. While Mean-Teacher model applies consistency regularization between teacher and student networks, FixMatch leverages high-confidence pseudolabel prediction and consistency constraints across different augmentations of the same sample. Both methods show improvements in spot localization by incorporating unlabeled data, achieving accuracies of 97.79% and 97.89%, respectively, with only 10% of data labeled. Furthermore, we evaluate position prediction using both 1-D multilayer perceptrons and 2-D convolutional neural networks under varying numbers of Wi-Fi receivers. The semisupervised Mean-Teacher model outperforms other supervised or semisupervised approaches, particularly with the 2-D convolutional neural networks (CNNs) and multiple receivers.