Magnetic Indoor Localization through CNN Regression and Rotation Invariance

A team of researchers from Fraunhofer FOKUS and TU Berlin has published a paper on arXiv demonstrating a novel approach to indoor positioning using convolutional neural networks (CNNs) and magnetic field data. Their models, MagNetS and MagNetXL, use rotation-invariant features derived from the 3D magnetic field—specifically the norm (Mn) and the projection onto the gravity axis (Mg)—to regress (x, y) positions directly. This solves a critical problem with raw 3D magnetometer data, which is highly sensitive to device orientation and degrades significantly under rotation.

Tested on the MagPie dataset covering three buildings (Loomis, Talbot, and CSL), the models showed that 2D (Mn, Mg) inputs maintain rotation-invariant accuracy and outperform raw 3D inputs once rotation exceeds building-specific thresholds: 0° for Loomis, 5° for Talbot, and 6° for CSL. MagNetXL achieves state-of-the-art accuracy, while MagNetS delivers similar performance with roughly one third of the parameters, making it suitable for mobile deployment. This work opens the door to infrastructure-free indoor navigation and IoT applications in GNSS-denied environments.