CompFi: Partially Connected Neural Network Using Complex CSI Data for Indoor Localization

Abstract

Many recent papers have directed attention to wireless indoor localization using Channel State Information (CSI) from the IEEE 802.11 OFDM scheme. Compared to the Received Signal Strength Index which contains only single source information, CSI from wireless communication contains channel characteristics per-subcarrier and thus brings higher positioning accuracy. Nevertheless, wireless interference, attenuation and multi-path problems mean that one cannot easily get exact location of a transmitter device. In this paper, we propose CompFi, an offline/online localization system for 5 GHz Wi-Fi that exploits both phase and amplitude of CSI complex values given in a phasor format as fingerprint data. Our novel Partially Connected Neural Network consists of 3-layer partially and 1-layer fully connected neural networks to make the best use of the CSI characteristics. Using regression analysis, our device-free localization system achieved 2-D distance error of about 1.74 m even under slight movement of the transmitter device at grid training and test points in a room environment with many structures.