Millimeter-wave (mmWave) and terahertz (THz) communication systems adopt large antenna arrays to ensure adequate receive signal power. However, adjusting the narrow beams of these antenna arrays typically incurs high beam training overhead that scales with the number of antennas. Recently proposed vision-aided beam prediction solutions, which utilize \textit{raw RGB images} captured at the basestation to predict the optimal beams, have shown initial promising results. However, they still have a considerable computational complexity, limiting their adoption in the real world. To address these challenges, this paper focuses on developing and comparing various approaches that extract lightweight semantic information from the visual data. The results show that the proposed solutions can significantly decrease the computational requirements while achieving similar beam prediction accuracy compared to the previously proposed vision-aided solutions.

翻译：毫米波（mmWave）和太赫兹（THz）通信系统采用大规模天线阵列来确保足够的接收信号功率。然而，调整这些天线阵列的窄波束通常会产生与天线数量成比例的高额波束训练开销。近期提出的视觉辅助波束预测方案，利用基站采集的\textit{原始RGB图像}来预测最优波束，已展现出初步的潜力。然而，这些方案仍具有显著的计算复杂度，限制了其在现实世界中的应用。为解决这些挑战，本文致力于开发并比较多种从视觉数据中提取轻量级语义信息的方法。结果表明，与先前提出的视觉辅助方案相比，所提方案在实现相似波束预测精度的同时，能够显著降低计算需求。