Abstract: Human trajectory prediction is an essential and challenging task in robot navigation and autonomous driving applications. One of the most challenging tasks is to model the interaction between pedestrians. Pairwise attention is used by most of the existing models to model the interaction. However, when there are too many pedestrians in the scene, these methods have redundancy in interaction modeling and ignore the interaction differences of pedestrians at different distances. To address these challenges, a multi-resolution global-local model is proposed, which contains a novel multi-region interaction sub-network to capture the global interaction and an additional local interaction sub-network to model pedestrians' interactions in the local neighborhood. In the meantime, the temporal attention mechanism is introduced in the proposed model to fuse the interactive information of different time steps. The experimental results show that compared with previous models, the proposed model achieves better performance on two publicly available datasets.

Key words: trajectory prediction, social model, motion planning