Machine-learning algorithm applied to magnetic localization

Abstract

Magnetic localization does not require line of sight, this makes it suitable for various applications, including indoor navigation, surgical tracking, motion capture, and 3D body scanning. Magnetic localization is typically an inverse problem, in which the magnetic field generated by several sources is measured at a given point, and the goal is to determine the coordinates of that point from these measurements. Multiple approach exists to perform such calculation. In this paper, the focus is put on machine learning algorithms, namely Random Forest and KNearest Neighbors. The results have been shown using simulations for the training of the algorithm and verified on experimental data. A sub-millimeter mean absolute error has been demonstrated on simulated data. A performance gap remains between simulated data and experimental one, partially due intrinsic errors of the machine learning algorithm, but also due to discrepancy between simulation and experiment. This work highlights the potential of machine learning in enhancing the precision and reliability of magnetic localization systems.