Robust Magnetic Fingerprint Positioning in Complex Indoor Environments Using Res-T-LSTM

Area of Science:

• Indoor positioning systems
• Geophysics and seismology
• Computer science

Background:

• Indoor location-based services face challenges due to WiFi limitations.
• Magnetic fingerprinting offers a complementary solution but struggles with environmental dynamics and sequence variations.
• Existing methods lack robustness against distortions in magnetic fingerprint sequences.

Purpose of the Study:

• To develop an advanced magnetic-fingerprint-based positioning model resilient to environmental dynamics.
• To improve the accuracy and stability of indoor positioning systems.
• To address challenges posed by sequence stretching, compression, and distortion in magnetic data.

Main Methods:

• Integration of Residual Networks (ResNet) for deep local feature extraction.
• Application of Transformer modules with self-attention for long-range dependency modeling.
• Utilization of Long Short-Term Memory (LSTM) networks to capture temporal dynamics.
• Proposed model named Res-T-LSTM combining these deep learning architectures.

Main Results:

• The Res-T-LSTM model demonstrated high performance across four distinct smartphone-carrying postures.
• Achieved a remarkably low average positioning error of 0.21 meters.
• Successfully mitigated issues related to sequence stretching, compression, and distortion.

Conclusions:

• The Res-T-LSTM model offers a robust and accurate solution for magnetic-fingerprint-based indoor positioning.
• The integration of ResNet, Transformer, and LSTM effectively handles complex indoor environments and dynamic conditions.
• This approach significantly enhances the reliability of indoor location-based services.