| Indoor location based service has become an important part of many applications,widely used in asset tracking,location-based network access,precision marketing,shopping mall and other scenarios.Global Positioning System(GPS)can provide highprecision location service with global coverage in outdoor environment,but it is difficult for satellite signals to penetrate buildings and work properly in indoor environment.In the indoor Wi-Fi,Bluetooth and other RF signal is mainly used to locate users or objects.Position estimation results can be obtained by analyzing and modeling the received signal intensity and channel state.The indoor positioning method based on RF fingerprint has gradually become the focus of research in the field of indoor positioning due to its high positioning accuracy and high reliability.However,existing indoor location methods still have many problems,including poor location accuracy,difficulty in mining the deep localization features hidden in a large number of fingerprints,and insufficient security and reliability of indoor location in crowdsourcing scenarios.Based on these problems,this paper conducts in-depth research and comprehensive experiments.The main work of this paper is as follows:Firstly,the localization feature mining of RF fingerprints is studied.In this paper,the concepts of location semantics and location marker words are proposed,and the traditional fingerprint sequence is defined as a finite sequence composed of marker words.A feature learning technique based on distributed representation vector and selfattention mechanism DAT is designed to mine the deep spatio-temporal localization features hidden in a large number of unlabeled RF fingerprints,and the vector space of localization marker words is visualized to verify the interpretability of the proposed feature extraction technique.Secondly,using the DAT localization feature extraction technology proposed in this paper,explore the spatio-temporal dependence hidden in the localization signal from the input data to estimate the location,including single point positioning model and trajectory prediction model.The single point positioning model DATLoc can predict a single position coordinate from the input RF signal sequence.The indoor trajectory prediction model DATSeq uses a self-attention-based seq2 seq model to predict indoor walking trajectory.Comprehensive experiments are carried out on the self-collected Bluetooth data set and the public Wi-Fi data set,and the average positioning errors reach 1.4m and 2.2m,respectively.The experiments verify the effectiveness of the proposed indoor positioning technology.Finally,the security indoor positioning system is studied.Typical crowdsourcing systems use smart phones as a platform to collect fingerprints,and the update stage and online positioning stage of fingerprint database are vulnerable to malicious attacks.This paper proposes to build a secure indoor positioning system,which can effectively identify malicious RF fingerprints to ensure the security of the fingerprint database update stage,and ensure the robustness of the indoor positioning system under the attack of moving valid RF beacons and installing malicious beacons. |
