Adaptive Multi-Agent Trajectory Prediction via Transferable Multi-Motion-Property Attention Network

Accurately predicting the trajectories of multiple agents at urban intersections is a critical and challenging task in intelligent transportation systems. However, existing theories face the following two dilemmas: 1) the insufficient representation of the spatiotemporal interactions between agents; and 2) the domain shifting problem due to the potential differences in spatial distribution across different domains. To address these issues, this paper proposes a transferable multi-motion-property attention linear inference network framework (T-MPAT-LIN) for cross-domain predicting the agents’ future trajectories at urban intersections without traffic signals. First, multi-motion-property attention network (MPAT) explicitly captures the spatial interaction information between agents by employing an appropriate aggregation mechanism that effectively processes various motion properties, such as relative coordinates, displacements, velocities, and accelerations. Additionally, a diagonalized linear inference network(LIN) is introduced as the temporal inference module to infer future trajectories. Second, a domain adaptive adversarial learning method, is proposed to mine cross-domain-invariant spatiotemporal features, which can effectively discriminate the data differences and align agents’ potential spatial distributions across different domains. Finally, the T-MPAT-LIN network is verified by multiple real-world intersection datasets, and the experimental results show that the T-MPAT-LIN network significantly improves the prediction accuracy and achieves superior performance compared with the state-of-the-art methods in various transfer tasks.