journal

Self-Attention-Enhanced PSW-LSTM for 3-D Indoor Pedestrian Positioning With Integrated Wi-Fi, Magnetometer, and Barometer Sensors

Zhanpeng Zhang, Jiale Wang, Ming Xia, Chuang Shi, Weisong Wen, Deyou Zhang

IEEE Transactions on Instrumentation and Measurement · May 2025

3-d-indoor-positioning multisource-sensors-fusion rapid-fingerprint-database-establishment self-attention-enhanced-presequence-warmup-long-short-term-memory(psw-lstm) smartphone

Abstract

The performance of smartphone-based fingerprinting indoor positioning methods heavily depends on the database quality and matching algorithms, typically providing only 2D coordinates. Traditional methods also face challenges like labor-intensive data collection, limited accuracy, and reduced longterm reliability. This paper proposes a Self-Attention-Enhanced Pre-Sequence Warmup Long Short-Term Memory (PSW-LSTM) framework, integrated with WiFi, barometer, and magnetometer sensors, to achieve accurate 3D pedestrian indoor positioning. Specifically, foot-mounted positioning devices are introduced to assist smartphones in collecting sparse fingerprint data, improving database efficiency. The proposed Self-Attention-Enhanced PSW-LSTM framework improves initial matching accuracy, reduces convergence time, and enhances trajectory continuity. Experimental results demonstrate that, in typical indoor environments such as educational and office spaces, the horizontal and vertical positioning accuracy for pedestrians is better than 2 meters and 0.5 meters, significantly improving smartphone-based 3D indoor positioning performance.

Problem Addressed

Smartphone-based fingerprinting indoor positioning depends heavily on database quality and matching algorithms, typically provides only 2-D coordinates, and faces labor-intensive data collection and degraded long-term reliability.

Key Contributions

Foot-mounted positioning devices assisting smartphones in sparse fingerprint collection for efficient database construction
Self-attention-enhanced PSW-LSTM framework fusing Wi-Fi, barometer, and magnetometer measurements
Improved initial matching accuracy, faster convergence, and enhanced trajectory continuity
Public release of the signal fingerprint database and testing data

Conclusions

In typical indoor environments such as educational and office spaces, the method achieves horizontal positioning accuracy better than 2 m and vertical accuracy better than 0.5 m. The model maintains strong stability even on datasets collected months apart, significantly improving smartphone-based 3-D indoor positioning performance.