Anchor-based Multi-modal Transformer Network for Pedestrian Trajectory and Intention Prediction

Abstract

Accurate prediction of pedestrian behavior is of paramount importance in enhancing both traffic efficiency and pedestrian safety for autonomous vehicles. However, due to the inherent unpredictability of pedestrian movements, achieving accurate predictions is challenging. In recent years, researchers have explored various approaches to address this challenge, focusing on multimodal trajectory prediction and crossing intention prediction. In this study, we propose an innovative method called the Anchor-based Multimodal Transformer Network for Pedestrian Trajectory and Intention Prediction. Our proposed approach adopts a multi-task learning framework that combines trajectory prediction and intention prediction, leading to improved performance in both tasks. Moreover, this integration enables a more comprehensive understanding of pedestrians' future movements, empowering autonomous vehicles to make more informed decisions. Additionally, to capture the diverse range of future possibilities for pedestrians, we introduce learnable intention anchors. Each anchor serves as a reference point and facilitates the prediction of a specific future modality. These anchors are iteratively refined by leveraging historical context in our multi-layer decoder, resulting in enhanced multimodal trajectory and intention predictions. Experimental result shows that our model achieves the state-of-the-art performance on JAAD dataset. Specifically, our method improves trajectory prediction by 9% and intention prediction by 11%. By providing comprehensive and accurate predictions of pedestrian behavior for automated vehicles, our model contributes to enhancing traffic efficiency and pedestrian safety.