Deep learning-based radar-assisted beam prediction

Beam selection in millimeter wave (mmWave) communication systems rely on information about the environment surrounding the communication target, and the use of deep learning methods to analyze sensing data acquired by low-cost radar sensors can effectively reduce communication overhead. In this paper, we further investigate the radar-based beam selection problem using deep learning methods. The beam selection performance of the Feature Pyramid Network (FPN) network and an optimized version of the Residual Networks (Resnet) network is evaluated for a large-scale real-world dataset, DeepSense 6G, and a targeted network is proposed for beam selection. The experimental results show that the accuracy of beam selection is improved by 18.5% compared to the original Lenet network.