PIEPredict++: An Improved Pedestrian Intention Estimation Model Incorporating Comprehensive Environment Information

Yikai Bao, Yukihiro Saito, Nobuhiko Nishio

Abstract: Predicting pedestrians’ highly variable crossing intentions has attracted considerable attention in autonomous driving research. This paper introduces PIEPredict++, a new model for predicting pedestrian intentions. It uses dynamic and static environmental features around pedestrians to infer intentions. To understand pedestrian behavior, we propose the EnvProcess module, incorporating the cross-attention mechanism to focus on key features in different environments. Evaluation of the publicly available pedestrian intention estimation dataset demonstrates that our model achieves exceptional results in terms of accuracy and F1 score. Our state-of-the-art performance in F1 score and recall is particularly noteworthy, underscoring the model’s exceptional ability to predict false-negative instances and its considerable implications for autonomous driving.

External IDs:dblp:conf/isvc/BaoSN24