Spatio-temporal Attention Transformer Model for Future Trajectory Forecast

Abstract

Human trajectory prediction is very important for people's safety. And human motion is essentially a multimodal motion, which brings great challenges to the trajectory prediction. Meanwhile, most sequence-to-sequence models only consider one of the temporal and spatial characteristics of pedestrian motion and ignore other. Therefore, based on the temporal and spatial attention mechanism and the transformer model, we fuse the temporal and spatial characteristics of the imputation, and propose a deep learning prediction model based on the spatial-temporal attention transformer mechanism (STATM). Our model designs a temporal attention module based on multi-head attention, a spatial attention module based on GAT, and finally maps spatial-temporal features to trajectory destinations through a transformer-decoder-like module. Compared with other methods, our model achieves better results on ETH and UCY public datasets.