Goal-Oriented Pedestrian Trajectory Prediction Considering Spatial–Temporal Interactions

Accurately predicting pedestrian trajectories in complex and dynamic scenarios is vital for the success of intelligent transportation systems (ITSs) and autonomous vehicles (AVs). Given the flexible and purpose-driven nature of pedestrian behavior, not only the interaction in time and space needs to be considered in dense scenes, but also the guidance and restraint effects of pedestrian moving goals need to be considered. However, previous approaches ignore the spatial–temporal interactions with neighboring pedestrians when estimating goals. Accordingly, a goal-oriented spatial–temporal (GOST) pedestrian trajectory prediction network is proposed. The GOST algorithm models temporal dependence and spatial interactivity separately. It then extracts spatial–temporal interaction features to estimate goals and predict trajectories. Based on this, a novel temporal pattern attention (TPA) mechanism based on the gated recurrent unit (GRU), named TPA-GRU, is introduced to capture the dependence of historical temporal information on a single pedestrian. In addition, a two-layer multihead graph attention network (GAT) is devised to effectively aggregate direct and indirect spatial interaction features. After fusing spatial–temporal interaction features, a conditional variational autoencoder (CVAE) is employed to estimate more reasonable social movement goals, thereby improving the accuracy of predicted trajectories. Experiments on ETH and UCY datasets validate the superior performance of our method, showing a reduction of 25.71% in average displacement error (ADE) and 18.64% in final displacement error (FDE).