Research On Indoor Localization Method Based On Deep Neural Networks

With WiFi localization is currently the mainstream indoor localization method,and fingerprint database construction is crucial to WiFi-based localization systems;however,for accuracy,this approach requires enough data to be sampled at many reference points,which consumes excessive effort and time.In this paper,we use device-free localization to collect CSI data at reference points,then we convert CSI data into CSI feature maps and extend the fingerprint database using the proposed Wavelet Transform-Feature Deep Convolutional Generative Adversarial Network(WTF-DCGAN)model.Then,the fingerprint database composed of the expanded amplitude feature maps are sent to the convolutional neural networks for image classification training.Finally,by adjusting the model and parameters of the convolutional neural networks,a high-accuracy feature maps classification model can be obtained and we can achieve indoor localizaiton.The use of WTF-DCGAN accelerates convergence during the training phase,and substantially increases the diversity of the CSI feature map.The extended fingerprint database both improves the accuracy of the indoor localization system and reduces the human effort involved in fingerprint database construction.The classification of feature maps by using the trained convolutional neural networks model improves the accuracy of indoor localizaiton.The specific research contents are as follows:1.The analysis verifies that the characteristics of the CSI data in the WiFi signal collected at the fixed receiving point(RP)in the indoor environment when the experimenter is standing at each reference point position have the characteristics representing the position and can be used for localization performance.The initial fingerprint database is used for localization by converting the CSI data representing the position feature collected by the experimenter at each reference point position into feature maps to represent the matching fingerprint of the location.Then we use the neural network model to expand the initial fingerprint database,which replaces the other papers to increase the labor and time cost to expand the fingerprint database.2.A new type of Wavelet Transform-Feature Deep Convolutional Generative Adversarial Networks is proposed,which is an improved model of Deep Convolutional Generative Adversarial Networks(DCGAN).The improvement of the model draws on Wasseratein Generative Adversarial Networks(WGAN)model.The superiority of the proposed WTF-DCGAN model is verified by analyzing and comparing theWTF-DCGAN model with the DCGAN model and the WGAN model for the diversity of the generated amplitude feature maps and the convergence of the model.3.Wavelet Transform-Feature Convolutional Neural Networks(WTF-CNN)model is proposed to train and classify the CSI feature maps in the fingerprint database to achieve the purpose of localization.By training the expanded fingerprint database through the WTF-CNN model,a high accuracy model that can be classified for the reference point location CSI feature maps can be obtained.Finally,the final localization result is obtained by taking the position geometric center of the first several high probability classification results.