Learning Social Constraints for Human Trajectory Prediction

Human trajectory prediction is essential for ensuring the safe navigation and scenario interaction for autonomous vehicles and mobile robots. Human trajectory is influenced not only by the human itself but also by the constraints of the surrounding objects. Hence, modeling the exact social constraints is necessary. In this paper, we classify the social constraints methods into three kinds, i.e., LSTM-based methods, transformer-based methods and GCN-based method, to predict human trajectories. In addition, we compare their performance on ETH/UCY and SDD, and the experimental results display the superior performance of the transformer-based methods.