Abstract : This paper explores techniques for beam prediction using multi-modal sensing data. Specifically, our aim is to develop a deep learning model that predicts the optimal beam using information collected from Camera, LiDAR, Radar, and GPS sensors. For this purpose, we propose ResNet-SE, which integrates a squeeze-and-excitation network with ResNet, and PIformer, a newly designed model that incorporates pooling-based attention and Inception mixer. Experimental results demonstrate a 22＼% improvement in prediction accuracy and a 38＼% reduction in training time compared to the state-of-the-art model.
Index terms : beam prediction, deep learning, Transformer, multimodal learning, wireless communications