Title: Enhancing Indoor Localization via CNN Regression and Rotation-Invariant Magnetic Features

Abstract

In environments where Global Navigation Satellite System (GNSS) signals are unavailable, indoor positioning serves as a critical technology for applications ranging from Internet of Things (IoT) networks to indoor navigation. By leveraging magnetic field characteristics alongside convolutional neural networks (CNNs), researchers have developed a cost-effective, infrastructure-independent method for achieving high-precision location tracking. Although magnetic fingerprinting is a viable strategy for indoor positioning, conventional models that process raw 3D magnetometer data suffer from significant performance drops due to their sensitivity to device orientation.

To mitigate this issue, this study introduces two rotation-invariant features extracted from the 3D magnetic field: the magnetic norm (Mn) and the projection along the gravity axis (Mg). We utilized a lightweight, 7-layer dilated CNN architecture, designated as MagNetS/XL, which takes magnetic sequences as input to directly predict (x, y) coordinates. The system’s performance was rigorously assessed using the MagPie dataset, which comprises handheld trajectories across three distinct buildings, under conditions involving both fixed and random rotations of training and testing data.

Our analysis reveals that models relying on raw 3D inputs (Mx, My, Mz) experience isotropic increases in error when subjected to fixed 90-degree rotations, with accuracy further declining as random rotation angles increase. Conversely, inputs based on the 2D features (Mn, Mg) preserve consistent accuracy regardless of orientation. These 2D inputs outperform 3D counterparts once rotational deviations surpass specific thresholds unique to each building: 0 degrees for the large Loomis building, 5 degrees for the medium-sized Talbot building, and 6 degrees for the small CSL building.

In terms of performance, MagNetXL matches or surpasses current state-of-the-art accuracy levels on the MagPie dataset. Furthermore, the MagNetS variant achieves comparable results while utilizing approximately one-third of the parameters, making it particularly suitable for mobile deployment. These findings demonstrate that the robustness provided by rotation-invariant inputs compensates for the reduction in input dimensionality, enabling reliable mapping and localization without the need for orientation alignment or additional infrastructure.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC