In this letter, we present a novel framework of 3D scene based beam selection for mmWave communications that relies only on the environmental data and deep learning techniques. Different from other out-of-band side-information aided communication strategies, the proposed one fully utilizes the environmental information, e.g., the shape, the position, and even the materials of the surrounding buildings/cars/trees that are obtained from 3D scene reconstruction. Specifically, we build the neural networks with the input as point cloud of the 3D scene and the output as the beam indices. Compared with the LIDAR aided technique, the reconstructed 3D scene here is achieved from multiple images taken offline from cameras and thus significantly lowers down the cost and makes itself applicable for small mobile terminals. Simulation results show that the proposed 3D scene based beam selection can outperform the LIDAR method in terms of accuracy.

中文翻译：

---

用于毫米波通信的基于 3D 场景的波束选择

在这封信中，我们提出了一种新的基于 3D 场景的毫米波通信波束选择框架，该框架仅依赖于环境数据和深度学习技术。与其他带外边信息辅助通信策略不同，所提出的一种充分利用环境信息，例如从3D场景中获得的周围建筑物/汽车/树木的形状、位置甚至材料重建。具体来说，我们构建神经网络，输入作为 3D 场景的点云，输出作为光束索引。与激光雷达辅助技术相比，这里重建的3D场景是从相机离线拍摄的多张图像中实现的，因此显着降低了成本，适用于小型移动终端。