WiFi Indoor Localization Algorithm Based On Deep Generative Model

With the widespread deployment and application of wireless AP(Access Point)in indoor environments,the technology of indoor positioning using WiFi has become a research hotspot.In recent years,indoor localization technology has had many outstanding works,however,few has been deployed in real-world scenarios.This is because the use of WiFi localization requires collecting a large number of WiFi RSS(Received Signal Strength)fingerprint data with location information at the reference points to form a fingerprint database.These fingerprint data are collected by professionals,and the acquisition process is time-consuming and effort-intensive.In addition,the localization accuracy depends heavily on the density of the reference points.The higher the reference points density,the higher the localization accuracy,which further increases the cost of high-precision indoor localization.How to reduce the collection cost of WiFi RSS fingerprint data and design a low-cost and high-accuracy WiFi RSS-based indoor localization algorithm is still a serious challenge.In this paper,a large number of low-cost and high-accuracy WiFi RSS indoor localization algorithm is realized by using a large number of fingerprint data(unlabeled fingerprints)without location information and a small amount of fingerprint data(labeled fingerprints)with location information.Among them,the leverage of large number of unlabeled fingerprints can greatly reduce the cost of fingerprint acquisition.Based on this,this paper proposes a new semi-supervised depth generation model---Semi-supervised Conditional Variational Auto-Encoder(SCVAE),which learns unlabeled fingerprints and labeled fingerprints.SCVAE is able to generate virtual WiFi RSS fingerprints at any given location by learning the underlying probability distribution of unlabeled fingerprints and labeled fingerprints.In addition,in order to enable SCVAE to be applied to large-scale scenarios(such as large-scale shopping malls),this paper proposes two effective training strategies,pre-training and joint training,to make the model parameters converge better.Based on the SCVAE model,the WiFi indoor localization system Deep Print(Deep Learning of finger Printing)is further designed and implemented on Android smartphones and servers.This paper evaluates Deep Print in three real-world scenarios.The experimental scenarios consists of two office building corridors and a large shopping mall with a total area of over 8,500 m~2.The result shows that Deep Print only needs 6 reference points(the interval between adjacent reference points is 60 meters)in the office buliding corridors,and the median localization error can reach 3.0 meters.In the shopping mall,compared with the classical localization algorithm RADAR and the newest Modellet,Deep Print only needs 16.5%of the reference points to achieve similar localization accuracy.In summary,Deep Print can effectively reduce WiFi RSS fingerprint acquisition costs.It is a low cost and precision indoor localization system,which can be deployed and applied in large-scale scenarios.