Computer Vision Aided Beamforming Fused With Limited Feedback

Abstract

The growing antenna array scale, the uncorrelated fadings between downlink and uplink of frequency division duplex (FDD) or analog beamforming design increases the difficulty of channel sounding or estimation. Non-wireless channel detection or beam weight prediction method is a promising solution to help obtain timely and accurate wireless channel state. Beamforming can be enhanced by the powerful sensing capability of cameras, for which this paper proposes a straightforward beam weight prediction method by implementing convolutional neural network (CNN) on images from cameras and a loss function based on chordal distance for this paper's task. Then a fusion method of visual detection and wireless sounding is developed to further improve spectral efficiency. This fusion method also utilizes a codeword rotation mechanism with Householder transform to save the notification overhead of visual detection results. A testbed has been built to verify the proposed approach with field measurement data. The proposed straightforward method is able to reach high spectral efficiency performance, and the fusion method could outperform exclusive visual detection or wireless sounding with appropriate hierarchical codebook. Beamforming, computer vision, hierarchical codebook. INDEX TERMS