Joint trajectory prediction in complex traffic scenarios using causal behavior analysis and risk potential fields

Accurate trajectory prediction is a cornerstone for ensuring the safety and operational reliability of autonomous driving systems. However, conventional data-driven models often exhibit significant performance degradation in complex, highly interactive traffic scenarios. Insufficient modeling of directed vehicle interactions and the lack of explicit constraints on predicted inter-agent proximity can lead to conflicting forecasts. To bridge this gap, this paper proposes a novel joint trajectory prediction framework integrated with causal interaction modeling and physics-informed risk constraints. First, a feature encoding module is developed to handle the heterogeneity of multi-agent interactions by decoupling spatiotemporal representations. Second, we develop a semantic causal interaction module that combines relative motion with directional relationships to model how neighboring vehicles influence target prediction, helping eliminate spurious spatial correlations. Finally, a Physics-Informed Risk Reconstruction Module jointly refines multi-agent trajectories using a Gaussian risk potential field and bounded gradient updates, reducing predicted proximity and improving motion regularity. Experimental evaluations on public benchmarks (Argoverse 2, TOD_VT, and nuScenes) demonstrate that the proposed framework significantly outperforms state-of-the-art methods, achieving error reductions of up to 5% in minFDE on Argoverse 2, 12% in Miss Rate on TOD_VT, and 13% in minFDE on nuScenes. These results demonstrate that this method offers significant advantages in improving trajectory prediction accuracy and safety in complex, highly interactive scenarios.