Framework for Autonomous Driving Trajectory Prediction via Physics-Constrained Data Fusion and Spatiotemporal Encoding

With the rapid development of artificial intelligence, autonomous driving has advanced significantly. Vehicle trajectory prediction, a key component of autonomous driving, directly affects the quality of planning and decision-making in dynamic traffic scenarios and is crucial for intelligent transportation systems. This paper proposes a collaborative prediction framework integrating physical and data-driven models. The physical model, enhanced by driving style parameters, better characterizes vehicle kinematics and driving styles for improved trajectory prediction. The data-driven model employs a spatiotemporal feature decoupling module with a dual-stream attention mechanism to extract and fuse temporal and spatial features from historical trajectories, enabling accurate future trajectory prediction. Through a bidirectional constraint framework, the two models achieve collaborative parameter optimization: the physical model provides kinematic feasibility constraints to ensure predictions align with real-world motion laws, while the data-driven model compensates for unmodeled dynamic factors, addressing the limitations of physical models in complex scenarios. Experimental results on the NGSIM and HighD data sets validate the superiority of the proposed method, providing a novel solution for trajectory prediction in autonomous driving.