WiFi Indoor Localization Algorithm Based On Generative Adversarial Networks

Many existing WiFi-based indoor positioning technologies can achieve high accuracy in laboratory environments.However,it is hard for them to be deployed in practical large-scale indoor environments,such as shopping malls and airports.One of the primary reasons is the high cost of fingerprint database construction.SCVAE is the state-of-the-art algorithm aiming to high-accuracy and low-cost localization by generating virtual fingerprints via deep learning.Unfortunately,it cannot generate high-accuracy virtual fingerprints at points far away from the reference points,which impacts its localization performance.In this dissertation,we propose Loc GAN(Localization via Generative Adversarial Networks)to achieve high accuracy yet low-cost indoor localization by leveraging a small number of labeled data and a large number of unlabeled data.Loc GAN is based on SCVAE.It consists of three stages:model training,fingerprint generation and online positioning.In the training phase,pre-training and adversarial training are used to improve the accuracy of the model.Pre-training is used to train the discriminator,regressor and CVAE independently before adversarial training.Adversarial training is used to confront the SCVAE generator and CNN discriminator by a small number of labeled fingerprints and a large number of unlabeled.The fingerprint generation module generates virtual fingerprints in the area of interests and then combine them with real fingerprints forming a hybrid fingerprint database.The online positioning module is a positioning algorithm that calculates the position of given fingerprints based on the hybrid fingerprint database.We evaluate the performance of Loc GAN in real-world large-scale indoor scenarios,with a total area of 9500m~2.In a shopping mall,the average accuracy is 6.15m at reference point density of 60 meters apart.Compared with that of SCVAE,localization error of Loc GAN decreases by 0.43 m,and the 90%cumulative error is reduced by 1.5m.In the teaching building,the average accuracy of Loc GAN is 3 m,which is 0.23 m less than that of SCVAE.In summary,Loc GAN achieves high accuracy using a small number of reference points,and is suitable for deployment in large-scale indoor environments to provide high-accuracy positioning service.