Data-Driven State Estimation for Vehicular Networks Using Spatial Proximity and Deep Learning

Abstract

Modern vehicular networks generate massive amounts of real-time data encompassing vehicle speed, GPS location, signal strength, packet delay, congestion levels, and available bandwidth, which are crucial for proactive traffic management and network reliability. Traditional traffic monitoring systems rely on manual or threshold-based approaches to estimate vehicle priority and network congestion, often leading to delayed decision-making, inaccurate predictions, and inefficient resource allocation. These limitations are exacerbated in high-mobility scenarios where traffic density fluctuates rapidly, and vehicles interact dynamically with roadside units (RSUs), making conventional techniques insufficient for ensuring timely and safe routing decisions. Motivated by the need for a robust, real-time predictive framework, this research implements a machine learning pipeline that leverages both Classification and Regression Trees (CART) techniques to assess priority levels and traffic density accurately. Existing algorithms, such as Support Vector Machines with CART (SVM-CART) and Random Forest (RF) with CART (RF-CART), provide baseline predictive capabilities; however, they struggle with non-linear relationships, feature interactions, and overfitting under heterogeneous traffic conditions. To overcome these limitations, a Hybrid Stacked Ensemble (HSE-CART) is proposed, combining RF and Hierarchical Soft Tree (HSTree) models with linear Regression (LinR) for Regression and logistic regression (LR) for Classification as a final estimator. This hybrid approach captures complex dependencies, balances bias-variance trade-offs, and enhances both classification and regression performance. Overall, the research establishes a scalable, interpretable, and effective solution for high-mobility vehicular networks, supporting safer and more efficient traffic management in real-time environments.