STAG: A novel interaction-aware path prediction method based on Spatio-Temporal Attention Graphs for connected automated vehicles

Abstract

Understanding social interactions between a vehicle and its surrounding agents enables effective path prediction, which is critical for the motion planning and safe navigation of automated vehicles. Several existing studies adopt recurrent neural networks which are powerful in modeling temporal interactions. Nevertheless, these models fail to deal with the complex spatio-temporal correlations among the traffic participants, especially for long-term predictions. On the other hand, recent spatio-temporal approaches based on Graph Neural Networks (GNN) have demonstrated considerable potential in modeling the social spatial and temporal interactions between the neighboring agents. These approaches, however, consider the influence of distinct agents on one another within a scene either fixed or symmetric, which may not be true in real-world scenarios. In this work, we introduce a novel graph-based system for vehicle path prediction, namely Spatio-Temporal Attention Graph (STAG). More specifically, STAG explicitly activates social interaction modeling and relational reasoning with a directed graph representation while considering spatial inter-agent correlations and motion tendency aspects. We assess the performance of STAG in three different driving scenarios, including highly structured highways, complex roundabouts, and highly dynamic unsignalized intersections. The results indicate that STAG outperforms several benchmark methods in terms of path prediction performance.