DoA Estimation via Sparse Bayesian Learning in a Non-Cooperative Mode Using a Single RF Link

Recent research has increasingly focused on utilizing radio frequency (RF) signals for indoor sensing. Traditionally, this involves employing antenna arrays configured across multiple RF links to capture key channel parameters, such as Direction of Arrival (DoA). However, this architecture requires each sensor to have an independent RF link, which increases complexity and cost. Additionally, deploying sensing systems necessitates pre-calibration of the RF links, further laboring the deployment. To address these challenges, we design a switched antenna array (SAA) that can time-division activate each antenna within the coherence time on a single RF link, thus simulating a multi-RF links platform for accurate DoA estimation. Subsequently, we develop a switching strategy and introduce a random forest-based matching algorithm to tackle the issue that signals from different antennas exported from the single RF link cannot be differentiated due to the non-cooperative mode between the SAA and receiver. Additionally, we compensate for the carrier frequency offset caused by asynchrony between transceivers, which affects DoA estimation in the SAA system. However, residuals still remain and can be considered as an enhancement to the noise. We introduce a DoA estimation algorithm based on sparse Bayesian learning that treats noise as its hyperparameter, enhancing the robustness against noise and improving the accuracy of DoA estimation. We build prototype systems and conduct field trials in real-world environments. The experimental results show that our system achieves 4.32° angle of arrival and 4.48° elevation of arrival median estimation errors by utilizing only a single RF link.

I. Introduction

Recently, with the advancement of integrated sensing and communication (ISAC) technology, widespread research focuses on developing sensing systems based on commercial-offthe-shelf (COTS) communication devices, such as WiFi and 5G base stations. These systems, including indoor localization [1], human tracking [2], and activity recognition [3], typically integrate channel parameters extracted from channel state information (CSI), such as the direction of arrival (DoA) [1], [2], into sophisticated pre-developed sensing models, complemented by intelligent optimization algorithms. Therefore, accurately acquiring parameters such as DoA is essential for implementing these systems [1], [2].