Occlusion-Aware Vision-Aided Beam Tracking for Multi-User V2I mmWave Networks

Predictive beam tracking is regarded as one of the most promising solutions to support vehicle-to-infrastructure (V2I) communication with requirements of low latency and high data rates. Especially in millimeter-wave (mmWave) V2I scenarios, vision-aided beam tracking based on artificial intelligence (AI) effectively mitigates critical delays caused by beam alignment. This study builds upon previous work on vision-aided beam tracking in mmWave V2I scenarios by enhancing alignment accuracy through occlusion-aware processing. We propose a visual semantic-aided predictive beam tracking (VS-PBT) method, utilizing a Sequence-to-Sequence (Seq2Seq) model and binary mask sequences for multi-frame beam prediction. Furthermore, using a deep learning-based approach, our method accurately detects the target vehicle in multi-user environments, ensuring precise beam alignment in complex scenarios. Simulation results show that the proposed VS-PBT achieves near 90% Top-1 beam tracking accuracy, achieving a 4% gain over single-frame prediction and exceeding modality benchmarks by over 20%. Even in visually occluded scenarios, VS-PBT maintains robust Top-$k$ beam tracking and near-optimal transmission rate ratio (TRR) performance.