Multi-Context enhanced Lane-Changing prediction using a heterogeneous Graph Neural Network

Abstract

Lane-changing Prediction (LCP) is crucial in defining vehicle movement in Microscopic Traffic Load Simulation (MTLS), impacting the distribution of traffic load on bridge decks. Despite their simplicity, existing physics-based approaches are subjective and deterministic, resulting in low fidelity in reflecting real-world scenarios. Current data-driven methods attempt to address this but only consider the trajectories of the subject vehicle and adjacent vehicles, neglecting other relevant contexts and thus compromising prediction accuracy. This study introduces LaneMCGNN, a multi-context enhanced graph neural network model for lane-changing prediction. The model integrates contextual features from spatial-temporal trajectories, vehicle types, and semantic maps, employing multi-attention mechanisms and Transformer modules to enhance feature extraction from these contexts. A lightweight Convolutional Neural Network (CNN) is utilized for efficient feature extraction from semantic maps of bridge decks. Trained and evaluated on an open-access dataset, our model achieves an accuracy of 98.928%, an F1-score of 0.989, and an Area Under Curve (AUC) of 0.999. Comparative discussions and ablation tests underscore the superiority of our model and the importance of incorporating multiple contexts. The proposed model can significantly enhance MTLS by improving the prediction of lane-keeping and lane-changing behaviors of vehicles, thereby increasing the precision of performance assessment for bridge components.