Behaviour-aware physics-informed trajectory prediction for freeway on-ramp merging vehicles

Abstract

Freeway merging areas involve intense interactions and non-stationary traffic dynamics, making accurate on-ramp trajectory prediction essential for safe and cooperative driving. However, existing deep learning methods often overfit and fail to maintain physical consistency under behavioural ambiguity. To address these issues, this study proposes STAI-eIDM, a behaviour-aware information fusion framework that integrates data-driven interaction learning with physical motion priors. Specifically, a Spatio-Temporal Attention-based Interaction network (STAI) captures multi-directional dependencies, while an enhanced two-dimensional Intelligent Driver Model (eIDM) provides physically interpretable kinematic constraints. A behaviour-aware gating mechanism further distinguishes yielding and non-yielding merging states. Experiments on the NGSIM I-80 dataset show that STAI-eIDM achieves a 5-second Root Mean Square Error (RMSE) of 4.09 m, reducing error by 11.47% over the data-driven baseline. Transfer results further indicate strong performance with only 5%–20% target-domain fine-tuning data. Overall, STAI-eIDM offers an accurate, interpretable solution for trajectory prediction in complex merging scenarios.