Modelling Diverse Interactions and Multimodality for Pedestrian Trajectory Prediction

Abstract

Pedestrian trajectory prediction can significantly enhance the perception and decision-making capabilities of autonomous driving systems and intelligent surveillance systems based on camera sensors by predicting the states and behavior intentions of surrounding pedestrians. However, existing trajectory prediction methods remain failing to effectively model the diverse and complex interactions in the real world, including pedestrian-pedestrian interactions and pedestrian-environment interactions. Besides, these methods are not effective in capturing and characterizing the multimodal property of future trajectories. To address these challenges above, we propose to devise a hand-designed graph convolution and spatial cross attention to dynamically capture the diverse spatial interactions between pedestrians. To effectively explore the impact of scenarios on pedestrian trajectory, we build a pedestrian map, which can reflect the scene constraints and pedestrian motion preferences. Meanwhile, we construct a trajectory multimodality-aware module to capture the different potential mode implicit in diverse social behaviors for pedestrian future trajectory uncertainty. Finally, we compared the proposed method with trajectory prediction baselines on commonly used public pedestrian benchmarks, demonstrating the superior performance of our approach.