Image Filtering Techniques for Beam Prediction in a Real-world 6G UAV Scenario

Millimeter-wave (mm-wave) and terahertz (THz) communication systems can satisfy the high data rate requirements in 5G, 6G, and beyond networks, but still rely on the use of extensive antenna arrays to guarantee sufficient received signal strength. Many antennas incur high beam training overhead; thus, the narrow beams require adjustment to support highly mobile applications. Deep learning (DL) vision-aided solutions can potentially forecast the optimal beams, leveraging raw RGB images captured at the base station. Image filtering techniques have been widely used in computer vision (CV) to modify and enhance the quality of an image, based on specific rules. This work applies different filters to RGB images for accurate mm-wave/THz beam prediction and feature extraction based on pre-trained convolutional neural networks (CNNs). The assessment of the developed framework is conducted on an actual dataset captured by an unmanned aerial vehicle (UAV) operating in the millimeter-wave (mm-wave) frequency range. The dataset comprises RGB images taken at the base station. Ensemble filtering techniques are also studied, enhancing the beam prediction accuracy of two state-of-the-art (SOTA) DL models.

I. Introduction

Vehicle-to-everything (V2X) communication has garnered significant research attention for both existing 5G and future 6G networks, catering to the needs of drones, unmanned aerial vehicles (UAV s), and autonomous vehicles [1]. To fulfill the high data rate demands of forthcoming aerial networks, unmanned aerial vehicles (UAVs) must be equipped with millimeter-wave (mm-Wave)/ terahertz (THz) transceivers and employ extensive antenna arrays. Also, by using narrow beams at both the transmitters and receivers, they ensure an adequate receive signal-to-noise ratio (SNR) [2].