Research On Indoor Localization Technology Based On Deep Learning

In recent years,with the progress and promotion of wireless communication,intelligent terminal and artificial intelligence technology,the demand for location-based services is growing rapidly.Because satellite navigation system can not provide high-precision location service indoors,indoor localization technology has important research significance and application value.With the characteristics of convenience and easy deployment,the new discovery of wireless localization and visual localization has become the key research direction of indoor localization technology.On the one hand,the channel state information overcomes many shortcomings of using only the received signal strength as the wireless localization feature.On the other hand,generating synthetic images from the building information model can greatly reduce the labor cost of sample acquisition in visual localization.Therefore,this thesis studies the integrated indoor localization technology based on channel state information and building information model.The specific work and innovations of this thesis are as follows:1.Aiming at the difficulty of Channel State Information(CSI)feature extraction,this thesis proposes a two-stage fingerprint localization method suitable for passive localization without equipment.Firstly,the CSI data is collected in the terminal,and then according to the CSI data of different subcarriers,receiving antennas,and multiple time spans,a localization fingerprint matrix with rich features is designed,and a localization fingerprint database is constructed.Then deep learning is introduced,and by adding random matrices to shortcut connections,shortcut connections are randomly established to enhance the integration ability of residual networks,and a neural network Res Fi that can efficiently extract CSI fingerprint localization features is designed.Finally,comparative experiments are carried out in a common indoor environment,and the results show that the Res Fi localization method is superior to other common localization methods.2.Aiming at the problem that CSI fingerprint features are not rich enough,this thesis proposes an improved passive localization method based on CSI phase and amplitude.Firstly,the linear transformation method is used to eliminate the measurement error in the phase to obtain stable phase information,and then the corrected phase and amplitude are fused to construct the CSI phaseamplitude fingerprint database.Considering the complex information in CSI amplitude-phase fingerprints,this thesis builds a localization network that can fully extract multi-scale features and key features in fingerprints based on the attention mechanism.To address the overfitting problem,the cross-entropy loss function is modified by using a label smoothing strategy.The experimental results show that the proposed improved method has better localization results.3.Aiming at the problem of insufficient feature completeness due to the limited information of a single localization source,this thesis proposes an indoor integrated localization method based on wireless signals and vision.The method uses the synthetic images rendered from the 3D building model and the CSI fingerprints of the corresponding positions as the training set,and then through the image reconstruction network and the discrimination network,the pose regression network is forced to learn the features without motion blur from the synthetic images with motion blur and reduce the impact of motion blur on localization.The experimental results verify the effectiveness and accuracy of the integrated localization method proposed in this thesis.