Vision and Position-Aware Beamforming via Attention-aided Convolutional Neural Networks

This paper designs an attention-aided deep learning module for beam prediction in millimeter-wave (mmWave) communication systems. The idea is to deploy a camera at the base station to obtain vision information for better beam prediction. Additionally, positional data obtained from the Global Positioning System (GPS) can be fused with image features to form a multi-modal input. Specifically, we combine a Convolutional Block Attention Module (CBAM) with the ResNet-18 backbone. As such, the neural network focuses on salient spatial and channel features and can achieve better performance in scenarios with multiple moving objects in the target area. The proposed method is evaluated using the DeepSense 6G dataset. Simulation results demonstrate that with the help of CBAM, our method outperforms existing baselines in beam prediction accuracy (increasing the Top-3 accuracy from 81.21% to 85.91%). Meanwhile, the computational complexity is only 0.01% higher than ResNet-18. Furthermore, it substantially reduces the beam search set compared to existing beam tracking methods, highlighting its practical utility in reducing signaling overhead in real-world deployments.