This paper proposes a flow matching (FM) framework based on rectified flow for vision-aided beam prediction in vehicle-to-infrastructure (V2I) links. Instead of modeling discrete beam index sequences, the method learns a continuous latent flow governed by an ordinary differential equation (ODE)-based vector field, enabling smooth beam trajectories and fast sampling. A terminal flow constraint enforces global consistency under finite-step integration, stabilizing long-term prediction. The resulting FM-based model significantly improves top-K accuracy over RNN and LSTM baselines, approaches the performance of large language model-based approaches, and achieves inference speedups on the order of 10 x and 10^4 x on identical GPU and CPU deployments, respectively.
Categories: eess.SP