Simplified Indoor Localization Data Acquisition by Use of Recurrent LSTM Networks on Sequential Geomagnetic Vectors

In order for indoor positioning services to be able to assert themselves across a broad front, simple processes with minimal costs and low user resistance are preferable. This is where our contribution kicks in: We present a method to detect individual positions within buildings by the use of locally induced distortions of the earth’s magnetic field with a smartphone alone, without any additional technology to locate the 2D position. To compensate for the lack of exact 2D coordinates, we fuse the sensor data into a gravitational magnetic vector (GMV) and perform a temporal classification based on a recurrent network. In this work we investigate the applicability of Long-Short-Term-Memory (LSTM) networks to find cross-correlations over a time frame. The trained models are available in a smartphone application. With this application the recognition rates of the locations are analyzed.