MMH-STA: A Macro-Micro-Hierarchical Spatio-Temporal Attention Method for Multi-Agent Trajectory Prediction in Unsignalized Roundabouts

Simultaneously predicting the future trajectories of heterogeneous multiple agents in a neighborhood is crucial for ensuring the safe and efficient operations of autonomous vehicles in intelligent transportation systems. The unsignalized roundabout is a typical traffic scenario wherein the multivariate and complex information interactions make the prediction task challenging. To address the aforementioned challenges, a macro-micro-hierarchical spatio-temporal attention (MMH-STA) architecture, which can effectively extract the temporal and spatial features of multiple agents based on the interaction mechanism, is presented in this article. This work makes three contributions: 1) A novel hierarchical framework, which considers the heterogeneity of different types of agents, is proposed for trajectory prediction in the roundabout environment. Similarly, the macro-state for interaction with the roundabout structure and the micro-state for interaction with other agents are introduced for an agent. 2) A heterogeneous graph is devised to represent the spatial interactions of a multi-agent, which is reflected in connections between different types of agents and the properties of nodes and edges in the graph. 3) A novel heterogeneous graph attention network with a multi-order neighborhood is designed to describe the spatial feature interactions in the neighborhood. Finally, a characterized decoder forecasts the future trajectories of multiple agents concomitantly. The experimental results reveal that the proposed model can effectively implement multi-agent trajectory prediction in roundabout scenarios with high accuracy and state-of-the-art performance compared to the baseline.