MINLOC:Magnetic Field Patterns-Based Indoor Localization Using Convolutional Neural Networks

Abstract

Conventional geomagnetic field-based indoor positioning and localization techniques determine the user's position by comparing the database with the geomagnetic field strength collected by the user. However, the magnetic field strength collected from various devices varies significantly. So, the greater the difference between the geomagnetic field strength stored in the database and user collected geomagnetic field strength is, the lower the degree of location accuracy will be. The diversity of smartphone makes it impossible to develop a single database which can work with all the smartphones in the same fashion. Intending to solve these problems, this paper proposes the use of geomagnetic field patterns called MP (Magnetic Pattern) with CNN (Convolutional Neural Networks) to perform indoor localization. The database is constructed using the MP that occurs at the points of measurement while the location is calculated using CNN which matches the user collected MP with the database. A voting mechanism is contrived to combine the predictions from several CNNs and the user's position is finally estimated. To evaluate the performance of the proposed technique, Samsung Galaxy S8 and LG G6 are used in two buildings with different experimental environments and path geometry. The proposed approach is tested by two male and two female users for analyzing the impact of user heights. Experiment results show promising results; furthermore, the comparison analysis with other magnetic indoor localization approaches demonstrate that the proposed approach outperforms them.