Improving LSTM-based Indoor Positioning via Simulation-Augmented Geomagnetic Field Dataset

Indoor positioning systems are now extensively used, enabling the precise localization of a user within a predefined space. The effectiveness of these services, particularly through the use of the geomagnetic field as feature, is notable, primarily due to its widespread natural availability, bypassing the need for dedicated infrastructure. Algorithmically, to establish a unique correlation between the data obtained by sensors at a specific location and the reference data used by the system, even in environments where the geomagnetic field exhibits minimal variations, a Recurrent Neural Network has been employed. This involves inputting the user's path into the network. To avoid costly dataset collection we employ a simulation-augmented approach. Furthermore, we combine the output of the proposed algorithm with inertial measurement unit data. Therefore, the proposed system attains an average error of around 0.4 meters in simulations and 1.6 meters in real-world scenarios.