Vision Image Aided Near-Field Beam Training for Internet of Vehicle Systems

In this paper, we develop a novel beam training scheme for extremely large-scale multiple-input-multiple-output (XL-MIMO) system by exploiting the visual image information. Different from the conventional beam training schemes that consumes a large number of in-band (time/frequency) resources, the proposed scheme only leverages the out-of-band (vision image) information, which can efficiently reduce the training overhead. Specifically, we proposed a vision image-aided beam training cascaded framework integrating YOLOv5 and ResNet18 networks, where the YOLOv5 uses the object detection technique to extract the size and location information of the mobile vehicles (MVs) and the ResNet18 based the extracted information infers the optimal beam index without occupying in-band overhead. The simulation results demonstrate that the proposed vision image-aided beam training scheme outperforms the benchmark scheme.