Multi-Source Low Redundancy Data-Aided Beam Prediction for V2I Communication

Millimeter wave (mmWave) communication requires huge beam training overhead, which is highly undesirable in vehicle-to-infrastructure (V2I) communication, due to the requirement for low latency and high reliability. Sensory information can be utilized to reduce the beam training overhead. However, single-type sensors have limitations that result in inadequate performance when used independently for assistance, while employing multi-type sensors sensors for assistance presents challenges such as data redundancy and a large volume of data. To tackle the aforementioned issues, we propose a multisource low redundancy data-aided beam prediction (MLRDBP) scheme. Specifically, we first extract and fuse features from LiDAR, GPS, and camera data. Then, we employ the principal component analysis (PCA) algorithm to reduce the dimensionality and redundancy of the fused features. Finally, we design a classification model based on a multi-layer perceptron (MLP), training it with the fused low redundancy features to predict the optimal beam direction for communication parties. The simulation results indicate that the proposed scheme achieves the accuracy exceeding 76.9% for top-1 beam prediction, offering satisfactory performance with lower data redundancy compared to single-sensor-aided schemes and other existing schemes.