Heterogeneous agents trajectory prediction with dynamic interaction relational reasoning

Abstract

Accurate trajectory prediction for different types of agents in complex environments is crucial for enabling safe navigation planning. However, current trajectory prediction methods usually ignore the fact that the behaviors of different types of traffic participants exhibit first-order discontinuities. For example, at intersections, the movement behavior can abruptly shift between stopping, going straight, turning right, or turning left, due to the frequent interactions occurring between the agents and the constraints imposed by traffic rules. Their behavior is directly affected by their interactions with the surrounding agents and the environment. To address these challenges, we propose a trajectory prediction method for heterogeneous agents with dynamic interaction relational reasoning. We utilize type-specific encoders to extract dynamic features of agents from their historical states. Interactions between heterogeneous agents are abstracted as heterogeneous graphs with directed edge features, then processed by a novel graph attention network with dynamic relational reasoning designed to extract spatio-temporal interaction features. To capture dynamic interactions, the graph is evolved into a topologically and representationally dynamic graph. Spatial interaction discontinuities are handled by reconstructing subgraphs for different agents with dynamic and changeable features. Furthermore, to reason about interactions, a two-element relational representation is proposed to obtain dynamic relational reasoning. Finally, we conduct a validation test of the proposed model utilizing real-world datasets. The experimental results from different aspects demonstrate that our method can effectively capture the dynamic interaction features between heterogeneous agents, realize the trajectory prediction of heterogeneous agents, and achieve state-of-the-art performance.