A spatio-temporal encoding-decoding model with embedded dynamic kinematic constraints for vehicle trajectory prediction in highway scenarios

Accurate vehicle trajectory prediction is fundamental to the development of intelligent transportation systems. Traditional physics-based models, however, often struggle with long-term forecasting due to their limited capacity to represent complex environmental dynamics. In contrast, data-driven approaches can improve predictive accuracy by integrating contextual and environmental information, yet they typically lack physical interpretability and may diverge from the intrinsic kinematic properties of vehicles. Recently, hybrid frameworks that integrate physical modelling with data-driven learning have attracted increasing attention. Nonetheless, most existing hybrid schemes treat the iterative processes of the physical and learning components as independent, where short-term outputs from the physical model are directly fused at the final stage of the data-driven predictor. This architectural design propagates accumulated physical-model errors, ultimately degrading overall prediction performance. Moreover, many data-driven modules insufficiently capture multi-vehicle interactions and underlying spatio-temporal dependencies. To address these challenges, this study proposes a Spatio-Temporal Encoding-Decoding model with Dynamic Kinematic Constraints (STE-DKC) for vehicle trajectory prediction. During decoding, the kinematic model is adaptively determined according to the inferred driving intent and imposed as a constraint at each prediction step, ensuring real-time physical consistency and mitigating cumulative trajectory errors. Historical motion sequences are encoded using a Peephole LSTM structure, while inter-vehicle dependencies are effectively captured through dilated convolutional social pooling enhanced by a Self-Attention mechanism. Extensive experiments on the NGSIM and HighD highway datasets demonstrate that the proposed framework consistently outperforms state-of-the-art approaches, particularly in high-speed traffic scenarios. The framework is specifically designed and validated for structured highway environments.   Highlights Introduces a decoding method with dynamic embedded kinematic constraints. Employs multi-head self-attention to capture global dependencies among vehicles. Models local vehicle interactions via convolutional social pooling. Superiority is verified through experiments on the NGSIM and HighD datasets.