Research On Multimodal Indoor Localization Method With Location Privacy Weak-dependent

The "last mile" problem of Location Based Service(LBS)has led most indoor positioning methods to focus on improving the accuracy,stability,persistence,and maintainability of positioning.However,the transmission and computation of strong position correlation data from smart mobile devices to localization servers exposes user privacy to risk.The dissertation attempts to answer the question whether weak position correlation feature can be used to design a location generation model with low risk of location privacy leakage and can be used for indoor localization systems with high robustness and localization accuracy.The main research elements of the dissertation can be summarized as follows.First,propose the recursive plot fingerprint feature representation method with weak position correlation data.To effectively protect user location privacy in fingerprint localization systems,the inertial sensor data is divided into weak position correlation data and strong position correlation data.Recurrence plots are introduced to characterize the temporal relationships of multimodal trajectory data from inertial measurement units on smart mobile devices,and three deep feature generation networks are designed to extract global location features from the recurrence plot-based weak position correlation data.Second,propose a deep generative learning network-based adaptive localization method for dynamic and static data.The method utilizes a location feature vector based on a variational self-encoder at the localization server and a predictor combining a long-andshort term memory network(LSTM)and a convolutional neural network(CNN)at the smart mobile device for location computation.The mobile device sends only weak position correlation accelerometer and gyroscope data to the server for location query,and the server generates global location features characterizing its location and returns them to the smart mobile device,reducing the possibility of user location exposure.Finally,the dissertation proposed the performance of CSLoc which is a localization system framework carefully on the open-source dataset.The results show that the average localization error of CSLoc with multimodal fusion is reduced by 28.00% compared to the traditional magnetic field single fingerprint localization system.In addition,the location prediction methods based on depth-generated feature extraction and LSTM networks possess significant advantages in handling dynamic data,which effectively improve the generalization ability and privacy protection of fingerprint localization methods.