Magnetic indoor positioning has attracted widespread attention due to its device-free nature and universality; however, existing methods suffer from limited feature extraction and scale heterogeneity caused by variations in pedestrian walking speed. To address these challenges, this study aims to propose a novel geomagnetic positioning framework that integrates a multi-dimensional geomagnetic database with a multi-scale LSTM-Transformer (MLT) architecture.
The proposed method constructs a five-dimensional geomagnetic fingerprint database (three-axis magnetic field, magnetic field strength and horizontal components) to enhance the feature discriminability. The MLT model is designed to capture local temporal dependencies and global spatial correlations, addressing the scale inconsistency issue without relying on Inertial Measurement Units (IMUs). The framework included an offline stage for database construction and an online stage for real-time positioning.
Experimental results on the public MagPIE data set and the self-collected IMUST and UCM data sets show that the proposed method achieves 90th percentile errors of 1.34 m on IMUST, 2.05 m on UCM and 1.20 / 1.69 / 2.58 m on MagPIE (CSL/Loomis/Talbot), outperforming existing state-of-the-art methods, including Dynamic Time Warping (DTW), Recurrent Neural Network (RNN), Hierarchical Long Short-Term Memory (HLSTM), Multi-scale Adaptive Indoor Localization (MAIL) and the multi-scale Transformer classification model proposed by Wang et al. (2024a). The multi-scale fusion strategy improves positioning accuracy by over 36% compared to the single-scale approach.
This paper presents two main contributions: a multi-dimensional geomagnetic database enriching spatial features; and a pure geomagnetic positioning method eliminating IMU dependency via multi-scale spatiotemporal modeling.
This work offers a cost-effective, smartphone-based indoor positioning solution requiring zero infrastructure investment. Achieving 1.2–2.6 m accuracy without IMUs, it enables scalable deployment in malls, airports and warehouses for asset tracking, navigation and location-based services. By leveraging existing geomagnetic fields, it eliminates hardware costs and maintenance, unlocking a $17B indoor positioning market with immediate commercialization potential and strong ROI for facility operators.
This study proposes a zero-infrastructure smartphone indoor positioning solution with 1.2–2.6 m accuracy, enabling low-cost deployment in malls, airports and hospitals. It enhances public safety and emergency response, significantly improves mobility for the visually impaired and advances smart city accessibility initiatives.
The originality of this research lies in constructing a five-dimensional geomagnetic fingerprint database and designing a multi-scale LSTM-Transformer fusion architecture that solves scale heterogeneity issues caused by walking speed variations without requiring IMUs. By innovatively integrating a denoising autoencoder with spatiotemporal modeling, the proposed method achieves 90th percentile errors of 1.20–2.58 m on the public MagPIE data set, with multi-scale fusion improving positioning accuracy by over 36%, thereby providing a new paradigm for pure geomagnetic positioning.
