Deep Learning Approach for Optimal Localization Using a mm-wave Sensor

Short-range indoor localization is one of the key necessities in automation industries and healthcare setups. With its increasing demand, the need for more precise positioning systems is rapidly increasing. Millimeter-wave (mm-wave) technology is emerging to enable highly precise localization performance. However, due to the limited availability of low-cost mm-wave sensors, it is challenging to accelerate research on real data. Furthermore, noise due to the hardware components of a sensor incurs perturbation in the received signal, which corrupts the estimation of range and the angle of arrival (AoA). Owing to the huge success of data-driven algorithms in solving regression problems, we propose a data-driven approach, which employs two deep learning (DL) based regression models i.e., dense neural network and convolutional neural network, and compare their performance with two machine learning based regression models, linear regression and support vector regression, to reduce errors in the estimate of AoA and range obtained via a mm-wave sensor. Our main goal is to optimize the localization measurements acquired from a low-cost mm-wave sensor for short-range applications. This will accelerate the development of proof of concepts and foster research on cost-effective mm-wave based indoor positioning systems. All experiments were conducted using over-the-air data collected with a mm-wave sensor, and the validity of the experiments was verified in unseen environments. The results obtained from our experimental evaluations, both for in-sample and out-of-sample testing, indicate improvements in the estimation of AoA and range with our proposed DL models. The improvements achieved were greater than 15% for AoA estimation and over 85% for range estimation compared to the baseline methods.