A Spatio-Temporal Graph Transformer Network for Multi-Pedestrain Trajectory Prediction

Jingfei Zhu, Zhichao Lian, Zhukai Jiang

Abstract: Trajectory prediction is widely used in many fields, e.g., in autonomous driving system and traffic surveillance system, good trajectory prediction algorithm can help prevent collisions or traffic jam. In practical scenarios, pedestrian trajectory is mainly affected by two factors: 1) Temporality. pedestrian’s movement is continuous, and there is a certain connection between the future and the history of the trajectory. 2) Spatiality. In the scenario of multiple pedestrians, one can be influenced by the people around and may be forced to change their movement path. In this paper, we propose a spatio-temporal graph transformer network and make the following contributions :1) A new decoder structure is proposed. 2) A new memory mechanism is used to improve the continuity of the temporality of the trajectory. 3) HuberLoss is used as the loss function of the network for the first time and shows good results. We demonstrate that our model effectively improves the accuracy of prediction by validation on five common datasets of pedestrian trajectory in ETH[9] and UCY[10].