Multi-Modal Sensing-Aided Beam Prediction using Poolformer for UAV Communications

In this paper, we propose a deep learning technique that uses both camera image data and GPS data to predict the optimal beam for efficient beamforming in UAV communication systems. Previous research has proposed single-modal beam prediction models that use either camera image data or GPS data individually. However, such methods have limitations due to their sensitivity to the measurement environment and outliers. To overcome this, we suggest a new technique based on Poolformer, a derivative model of the transformer, which combines these two types of data. Experimental results show that the proposed Poolformer-based model outperforms the existing model in terms of Top-1, 2, 3 accuracy for both 32 and 64 beams. Notably, the Top-3 accuracy of the proposed model approached nearly 100% accuracy in both experiments.