Deep Learning-Based Vehicular Millimeter-Wave Channel Prediction Using Visual Information

This is the authors' abstract. We don't add key points for this paper.

Millimeter-wave channel prediction requires detailed and accurate environmental information due to challenges of non-line-of-sight transmission and severe reflection and scattering. Image information conveniently reflects real-time changes of physical environment and provides a promising solution for millimeter-wave channel modeling. However, research on millimeter-wave channel prediction based on RGB images is still in its exploratory stage. Meanwhile, RGB images introduce new influencing factors, such as visual angle changes, leading to inaccurate prediction. This article proposes a millimeter-wave channel prediction model based on propagation knowledge and visual information, which enhances original RGB image features while combining manual features from specific domain prior knowledge in propagation. Experimental results show that the proposed model outperforms traditional RGB image-based prediction for a multi-scenario, multiview fused dataset, achieving a symmetric mean absolute percentage error as low as 4.16%. In addition, the proposed model demonstrates enhanced stability and interpretability.