GCHGAT: pedestrian trajectory prediction using group constrained hierarchical graph attention networks

Predicting the motion of pedestrians is a challenge due to the uncertainty of the target pedestrian itself and the influence of other people in the environment. Modelling social interactions is of great significance for pedestrian trajectory prediction. However, most of the existing works only focus on the pair-wise interactions of humans but ignore the group-wise interactions. This paper proposes a group constrained hierarchical graph attention network, GCHGAT, to capture the intragroup, outgroup, and intergroup interaction separately. We first get a rough prediction via a vanilla generative adversarial network. Then, a state-refinement module is used to refine the rough prediction based on interaction information. We compare the performance of our method with related methods on the ETH and UCY datasets. The results show that our approach outperforms all benchmarks with the lowest average prediction error and successfully predicts complex social behaviours.