Learning Spatiotemporal Features of CSI for Indoor Localization With Dual-Stream 3D Convolutional Neural Networks

Recently, the research of WiFi-based indoor localization combing with deep-learning techniques has earned wide attention due to its potential applications in smart cities. In this paper, a novel fingerprinting system is proposed to achieve indoor localization via learning spatiotemporal features from channel state information (CSI) of multiple-input multiple-output wireless channels (CSI-MIMO) by a dual-stream three-dimensional (3D) convolutional neural network (DS-3DCNN). In the proposed system, the gathered raw CSI-MIMO data are firstly preprocessed through amplitude outliers elimination and phase sanitization for constructing a pair of 3D CSI-MIMO matrices including a 3D amplitude matrix and a 3D phase matrix. Next, the 3D matrices will be input to the DS-3DCNN deep neural network which consists of two parallel subnetworks with specific architecture of several convolution, batch normalization, max-pooling, and fully connected layers. Through this DS-3DCNN network, learning spatiotemporal features of CSI-MIMO is carried out simultaneously from 3D amplitude and phase matrices. And then, probabilistic classification results of two subnetworks are fused in the final output layer of the proposed DS-3DCNN based on Bayes' theorem. Moreover, in the offline training stage, a dual-stream joint optimization method is presented for efficiently optimizing network parameters. After offline training of the DS-3DCNN, in the online locating stage, current CSI-MIMO data are firstly collected from the mobile device to be located. Then probabilistic classification results are obtained from the output layer of the DS-3DCNN, and further used to approximate the posterior distribution of the mobile device's location with a Gaussian mixture model. Finally, a novel location estimation algorithm is deduced based on the minimum mean square error (MMSE) criterion. Since unique features of wireless MIMO channels are jointly learnt in spatial, temporal, and frequency domains, the proposed DS-3DCNN based fingerprinting system is reasonable to provide accurate localization results in indoor environments, which is verified in corresponding experiments.