Energy-optimization-based agent motion model: a hybrid approach integrating data-driven methods and trajectory prediction for intersection traffic simulation

Traffic intersections are currently among the most complex traffic scenarios, where many simulation methods underperform. Due to the lack of predefined lanes, agents have a high degree of freedom in the central area of intersections, making it difficult to guide their movement. Traditional data-driven models highly depend on the quality of input data, require large amounts of historical trajectories, and lack a unified method for providing reasonable decision-making for agents at each frame. Therefore, this paper proposes a new method that combines data-driven and trajectory prediction. Our method integrates an data-driven model and a trajectory prediction model, utilizing real data inputs while supplementing data with trajectory prediction, and guides agent motion using energy optimization equations. It fully considers multiple energy terms such as speed continuity, direction guidance, visual perception, and emergency handling to ensure that agents make reasonable decisions in realistic traffic scenarios. This simulation framework can realistically model the behavior of multiple agents in intersection scenarios. The experiments demonstrate the proposed method's effectiveness.