Social Transformer: A Pedestrian Trajectory Prediction Method based on Social Feature Processing Using Transformer

Abstract

In pedestrian trajectory prediction, the prediction accuracy depends largely on the consideration of the impact of social relations on the prediction object. Social pooling and graph neural networks (GNN) are two traditional social feature processing methods, they process sparse and nonuniform social features into more intensive and uniform information. In this paper, the Social Transformer Network (STNet) was proposed based on the GNN, which is a graph attention network. After a conditional variational auto-encoder (CVAE)-based preprocessing network provided a destination prediction, a transformer network was used to process the social feature data of the past trajectory and destination information. The transformer network was based on the self-attention mechanism, and it can assign different attention weights to different social features so that more attention is paid to the social relations with greater impacts on the pedestrian's trajectory. In this paper, STNet was tested on the ETH/UCY datasets. The results showed that average displacement error (ADE) was reduced by 17.2% and final displacement error (FDE) was reduced by 14.6%, indicating that the STNet improved the prediction accuracy.