STGT: Forecasting Pedestrian Motion Using Spatio-Temporal Graph Transformer

Abstract

Full understanding human motion is essential for autonomous agents such as self-driving vehicles and social robots for navigating in dense crowded environments. In this paper, we present a trajectory prediction framework which models inter-pedestrian behaviour through graph representations and then apply attention through a Transformer network to better forecast human motion. Previous works have incorporated pedestrian interaction using social and graph pooling mechanisms whereas our work utilizes complete graph structure of pedestrians which helps to obtain robust spatiotemporal representations. We also leverage semantic segmentation architecture to encode scene context. Our experiments highlight the potential of handing pedestrian interaction with graph convolutional networks and Transformer and, on top of that, shows marginal improvement with inclusion of semantic scene features.