Device-Free Localization Using Commercial UWB Transceivers

TL;DR

This paper introduces a deep learning-assisted particle filter method for device-free localization using commercial UWB transceivers, achieving high accuracy in real-world environments with low SNR and clutter.

Contribution

It presents a novel combination of deep learning and particle filtering to improve UWB-based device-free localization accuracy and robustness in challenging conditions.

Findings

Achieved about 15 cm RMSE in localization accuracy.

Demonstrated real-time processing with 4 ms average time.

Outperformed existing state-of-the-art methods in accuracy.

Abstract

Recently, commercial ultra-wideband (UWB) transceivers have enabled not only measuring device-to-device distance but also tracking the position of a pedestrian who does not carry a UWB device. UWB-based device-free localization that does not require dedicated radar equipment is compatible with existing anchor infrastructure and can be reused to reduce hardware deployment costs. However, it is difficult to estimate the target's position accurately in real-world scenarios due to the low signal-to-noise ratio (SNR) and the cluttered environment. In this paper, we propose a deep learning (DL)-assisted particle filter to overcome these challenges. First, the channel impulse response (CIR) variance is analyzed to capture the variability induced by the target's movement. Then, a DL-based one-dimensional attention U-Net is used to extract only the reflection components caused by the target and suppress the noise components within the CIR variance profile. Finally, multiple preprocessed CIR variance profiles are used as input to a particle filter to estimate the target's position. Experimental results demonstrate that the proposed system is a practical and cost-effective solution for IoT and automotive applications with a root mean square error (RMSE) of about 15 cm and an average processing time of 4 ms. Furthermore, comparisons with existing state-of-the-art methods show that the proposed method provides the best performance with reasonable computational costs.