PosFormer: Generalizable Indoor Positioning via Global--Local Feature Fusion Network

Area of Science:

• Robotics and Automation
• Wireless Communication and Networking
• Signal Processing

Background:

• Ultra-wideband (UWB) technology offers high-accuracy indoor positioning but struggles in multipath Non-Line-of-Sight (NLOS) environments.
• Existing methods, including geometry-based and deep learning approaches, face limitations in accurately capturing UWB signal propagation characteristics.
• Challenges include distance overestimation in traditional methods and insufficient feature extraction in current deep learning models.

Purpose of the Study:

• To develop an advanced deep learning model for high-accuracy indoor positioning using UWB signals.
• To address the limitations of existing methods in complex, multipath-rich environments.
• To improve the robustness and deployment efficiency of UWB localization systems.

Main Methods:

• Proposed PosFormer, a dual-fusion network combining Transformer and Convolutional Neural Network (CNN) modules to process Channel Impulse Responses (CIRs).
• Incorporated multipath physical information to enhance feature extraction.
• Introduced a nonadjacent anchor-subsets (NAASs) scheme for CIR diversity and a lightweight transfer learning (TL) framework for cross-environment deployment.
• Utilized public industrial datasets for extensive experimental validation.

Main Results:

• PosFormer achieved a Mean Absolute Error (MAE) of 17.64 cm, significantly outperforming baseline models (CNN, LSTM, Transformer).
• In challenging industrial hall environments, the TL framework enabled a pretrained model to reach 35.92 cm accuracy with only 20% of fingerprint data.
• Demonstrated superior performance in extracting both global and local multiscale features from UWB CIRs.

Conclusions:

• The proposed PosFormer model effectively addresses UWB indoor positioning challenges in NLOS environments.
• The dual-fusion network architecture and TL framework enhance accuracy, robustness, and data efficiency.
• PosFormer shows significant practical value for real-world UWB localization applications, especially in industrial settings.