A Deep Neural Network-Based Indoor Positioning Method using Channel State Information

Abstract

The channel state information (CSI) measurement, which characterizes the multipath channel between the transmitter and the receiver, can serve as a fingerprint for the receiver position. Recently, a series of deep learning-based indoor positioning fingerprinting (FP) methods using CSI have been proposed to enhance the localization performance. In this paper, we present a deep neural network (DNN)-based indoor positioning FP system using CSI, which is termed DNNFi. The proposed DNNFi, which maintains a single DNN instead of multiple deep autoencoders at different reference points, allows a faster computation for the online inference and a lower memory usage for the weights/biases. A stack of autoencoders is utilized to pre-train the weights layer-by-layer. The softmax function is utilized to decide the probabilities of the receiver position being on these reference points, which can be used to estimate the receiver position. Experimental results are presented to confirm that DNNFi can effectively reduce location error compared with the conventional CSI positioning FP approaches.