Sensor Fused Probabilistic Graph Forecasting for Persistent Traffic Groups Across Urban Networks

Abstract

The rapid urbanization of modern metropolitan areas has introduced unprecedented challenges in traffic management and autonomous vehicle navigation. Central to solving these challenges is the accurate identification and prediction of persistent traffic groups, which are clusters of vehicles that maintain spatial and temporal proximity over extended periods. Traditional single-sensor and deterministic models often fail to capture the complex, dynamic, and uncertain nature of these vehicular formations, particularly in environments characterized by occlusion, adverse weather, or erratic driver behavior. This paper introduces a comprehensive framework termed Sensor-Fused Probabilistic Graph Forecasting, designed specifically for the robust detection and future state prediction of persistent traffic groups. By integrating heterogeneous data streams from cameras, light detection and ranging systems, and radar sensors, the proposed methodology constructs a dynamic probabilistic graph. In this graph, nodes represent individual vehicles or spatial entities, while edges denote the statistical probability of interaction and group cohesion. Our approach leverages advanced graph forecasting techniques to propagate these probabilities over time, generating highly accurate, uncertainty-aware predictions of group trajectories and stability. Extensive evaluations on complex traffic datasets demonstrate that the proposed sensor-fused probabilistic framework significantly outperforms existing deterministic baselines in both detection accuracy and forecasting reliability, offering a critical advancement for intelligent transportation systems and automated driving safety protocols.