Machine Learning Driven Method for Indoor Positioning Using Inertial Measurement Unit

Abstract

The application of inertial measurement unit (IMU) is widespread in many domains, but the main hindrance in localization is the errors accumulation in the integration process over a long time. Recently, we notice that many researchers have applied machine learning (ML) algorithms to indoor positioning by using IMU sensor data, which sufficiently proves that the 6-dim data collected by IMU sensor contain a lot of information. In this paper, we present a ML driven method to make a regression between IMU sensor data and 2-D coordinates. To build a regression model with better generalization and lower computational complexity, this paper carries out feature extraction in the time-and time-frequency domain. The simulation run on Intel core i5-4200h shows that the method is able to suppress the drift of the inertial navigation system after a long-time travel. In comparison of GPS+IMU using extended Kalman filtering (EKF), the positioning RMS of our method on circular trajectories with a radius of 7 meters and 10.5 meters is reduced by at most 70.1% and 86.1%, respectively.