A Hybrid Machine Learning Approach for Scalable and Uncertainty-Aware RSU Tracking in VANET

This study addresses the challenge of real-time Roadside Unit (RSU) tracking in Vehicle-to-Vehicle (V2V) networks, where existing methods struggle with accuracy, scalability (100–25,000 vehicles), and noisy data. Traditional approaches lack adaptive feature learning and uncertainty handling, leading to misclassification in dynamic traffic conditions. To overcome these limitations, we propose a hybrid machine learning framework combining XGBoost (for high-accuracy classification) and a Gradient Discriminant Filter (GDF) (for probabilistic refinement), enhanced by Bayesian Deep Neural Network (BDN)-based feature transformation. The system processes real-time simulated vehicle tracking data, optimizing feature representations for improved interpretability and performance. XGBoost handles nonlinear decision boundaries, while GDF refines predictions under uncertainty. BDN modeling further enhances robustness by quantifying prediction confidence. Experimental results demonstrate over 95% classification accuracy with efficient scalability across varying traffic densities. The proposed method outperforms conventional techniques in both precision and computational efficiency, making it suitable for large-scale intelligent transportation systems (ITS). This work advances real-time V2V analytics by integrating ensemble learning, probabilistic filtering, and deep feature transformation for reliable RSU tracking in dynamic environments.