부산대학교 도서관

Predicting pedestrian crossing trajectories has become a primary task in aiding autonomous vehicles to assess risks in pedestrian–vehicle interactions. As agile participants with changeable behavior, pedestrians are often capable of choosing from multiple possible crossing trajectories. Current research lacks the ability to predict multimodal trajectories with interpretability, and it also struggles to capture low-probability trajectories effectively. Addressing this gap, this article proposes a multimodal trajectory prediction model that operates by first estimating potential motion trends to prompt the generation of corresponding trajectories. It encompasses three sequential stages. First, pedestrian motion characteristics are analyzed, and prior knowledge of pedestrian motion states is obtained using the Gaussian mixture clustering method. Second, a long short-term memory model is employed to predict future pedestrian motion states, utilizing the acquired prior knowledge as input. Finally, the predicted motion states are discretized into various potential motion patterns, which are then introduced as prompts to the Spatio-Temporal Graph Transformer model for trajectory prediction. Experimental results on the Euro-PVI and BPI datasets demonstrate that the proposed model achieves cutting-edge performance in predicting pedestrian crossing trajectories. Notably, it significantly enhances the diversity, accuracy, and interpretability of pedestrian crossing trajectory predictions.