HEI-GAN: A Human-Environment Interaction Based GAN for Multimodal Human Trajectory Prediction

Human trajectory prediction is an indispensable key component in autonomous driving systems and robot systems. The difficulty of human motion prediction lies in its inherent stochasticity and multimodality. Recently, some studies model the multimodality of human motion by predicting multiple possible future goals, which improves the accuracy of trajectory prediction. However, such methods fail to fully consider the influence of the environment on the target motion to be predicted. To solve the problem above, we propose HEI-GAN, a goal prediction model, with a human-environment interaction modeling method. The proposed method fuses the information from both the human motion and the environment to guide the model to learn the impact of human-environment interaction. We tested the performance of HEI-GAN on Stanford Drone Dataset and ETH/UCY dataset. The results show that HEI-GAN has a better prediction performance than the existing goal prediction models. At the same time, the prediction accuracy of the proposed method for the complete trajectory also reaches the advanced level.