Improved CNN-based Magnetic Indoor Positioning System using Attention Mechanism

Geomagnetic field fingerprinting is emerging as the most feasible and cost-effective alternative to WiFi and Bluetooth fingerprinting due to the omnipresence of magnetic field. With the increasing interest in applying deep learning methods to magnetic fingerprinting, attention mechanisms have been lately adopted to better recognize magnetic sequence patterns. In the last couple of years, several comparative studies have proved the added value of leveraging attention mechanisms in conjunction with Recurrent Neural Networks (RNNs) for indoor positioning systems treating sequential magnetic data. Yet, no study has been conducted to determine the performance contribution of these mechanisms to solutions based on Convolutional Neural Networks (CNNs). In this study, we propose a CNN-based indoor magnetic-only fingerprinting system using Recurrence Plots (RPs) as sequence fingerprints and approaching the localization problem from a regression perspective. Real-world data in an indoor environment are used to build the system, and fairly evaluate it before and after the introduction of an attention module. Our findings show that the architecture involving attention mechanism clearly outperforms the initial RP-based CNN, yet results in much higher prediction latency.