应用于毫米波车车通信的多模态感知辅助波束预测

Multi-modality Sensing Aided Beam Prediction for mmWave V2V Communications

为了确保车联网通信的传输可靠性,大规模多天线技术的毫米波通信亟需精确的波束赋形.在高动态的车辆通信环境下,传统的波束对准方式会造成巨大的资源开销,难以在相干时间内建立可靠链路.因此,文中提出基于多模态感知信息辅助的波束预测方案.该方案融合视觉和激光雷达点云两种非射频感知信息,利用深度神经网络进行多模态信息的特征提取,通过透视投影实现图像空间语义信息和物理空间位置信息的精准匹配和深度融合.通过协同感知坐标校正和车辆位置预测,将物理环境的特征精确映射到角域信道,从而实现实时准确的波束预测.在多模态感知仿真数据集(M3SC)上的测试结果表明,文中方案能实现较高的角度追踪精度和可达通信速率.

To ensure the transmission reliability of vehicular communication network,precisely aligned beamforming of millimeter-wave communication using massive multi-input multi-output(mMIMO)technology is urgently required.In highly dynamic vehicular communication scenarios,traditional beam alignment schemes incur significant resource overhead and struggle to establish reliable links within the coherence time.To address this critical challenge,a scheme of multi-modality sensing aided beam prediction for mmWave V2V communications is proposed.Two non-RF sensing modalities,vision and ranging(LiDAR)point cloud,are integrated,and deep neural networks are employed for feature extraction and integration of multi-modal information.Accurate matching and deep fusion of image space semantic information and physical space location information are achieved through perspective projection.By collaborative sensing coordinate calibration and vehicle position prediction,the features of physical environment are accurately mapped to the angular-domain channel,enabling real-time and precise beam prediction.The experimental results on the mixed multi-modal sensing-communication dataset(M 3SC)show that the proposed scheme achieves high angle tracking accuracy and achievable communication rate.