Spatio-Temporal-Interaction Graph Neural Networks for Multi-Agent Trajectory Prediction

Abstract For intelligent transportation systems, accurately forecasting the future trajectories of multiple agents is pivotal. Considering the increased diversity of agents within a scene, in order to capture and model the variations in their appearance, motion status, behavioral patterns, and interrelationships, we propose a simple yet effective framework based on Spatio-Temporal-Interaction Graph Neural Networks. Specifically, a Multi-Class Agent Encoder is meticulously tailored to the specific class of each agent to distill pertinent information from their motion attributes and historical trajectories. Subsequently, a Spatio-Temporal-Interaction Graph Attention Module is constructed to productively represent and learn the complex, dynamic interactions. Finally, a Multimodal Trajectory Generation Module is customized, and a learnable diversity sampling function is introduced to map the features of each agent to a set of potential variables, so as to capture the multimodal distribution of future trajectories. Empirical evaluations on the ETH/UCY and KITTI datasets reveal that our method can efficiently improve the accuracy of trajectory prediction.