Camera-Assisted Beam Selection via Unified Learning in 6G Vehicle-to-Infrastructure (V2I) Scenarios

The integration of non-RF data into communication-related processes, such as beam selection, can significantly enhance decision-making efficiency in scenarios requiring exhaustive search among numerous beam candidates. Recent studies have demonstrated the benefits of leveraging sensory data from cameras at base stations (BSs) to improve millimeter-wave (mmWave) and sub-terahertz (sub-THz) beam selection, helping to reduce reliance on beam sweeping and better support mobile applications. In this work, we propose a unified machine learning framework based on eXtreme Gradient Boosting (XGBoost) for beam selection in both single- and multi-candidate scenarios. Our method extracts object features using YOLOv4 and estimates the angle and distance of the target vehicle. By fusing visual and positional data, the model effectively predicts the optimal beam direction across diverse scenarios. In contrast to the baseline approach - which uses two separate models for single- and multi-candidate cases - our single unified network handles all scenario types without requiring architectural changes. We validate our model on six diverse real-world scenarios from the large-scale DeepSense 6G dataset, demonstrating strong performance even in combined datasets. This work highlights the practicality of vision-assisted beam prediction using a single model, paving the way for more adaptive and efficient wireless communication systems.