Towards more accurate mobile direction finding with UWB

Abstract

UWB is becoming increasingly more available to the general public as part of consumer devices such as smartphones and smartwatches. This opens opportunities for new indoor positioning paradigms because UWB in consumer devices not only supports ranging but also AoA (Angle of Arrival) estimation, meaning dependence on additional positioning infrastructure can be reduced. Since this is a recent development, however, not much research has been conducted on evaluating UWB performance in these consumer devices and how to improve it for better indoor positioning accuracy. To contribute to this research gap, this paper is the first to propose a machine learning solution to AoA accuracy improvement in UWB-equipped iPhones when communicating with DWM3001CDK sensors while in motion. The distinguishing feature of our solution is that, unlike previous works, it uses AoA measurements for training instead of raw CIR (Channel Impulse Response) data, meaning the anchors do not need to be attached to a computer for data collection, which makes the installation of anchors more convenient. In addition, our solution combines machine learning with a collaborative approach based on our positioning vector framework, which further improves AoA error. We compiled a training dataset based on real UWB measurements collected in a large indoor environment. Extensive experiments were conducted to evaluate different machine learning models, and our results show that machine learning can improve the 90th percentile AoA error from about 60° to 11° and thus improve the average direction estimation accuracy to 96.85%.