Map-enhanced generative adversarial trajectory prediction method for automated vehicles

Abstract

Trajectory prediction in dynamic and highly interactive scenarios is a critical method for achieving advanced autonomous driving. Maximizing the guidance and constraints provided by high-definition (HD) maps can help improve prediction performance across the board. In this paper, we propose a map-enhanced generative adversarial network (ME-GAN) for vehicle trajectory prediction. The vehicle motion features, map constraints, traffic flow density, and vehicle interactions are comprehensively considered in the generator, and a graph query mechanism is proposed to realize the reuse of the global map. In the discriminator, in addition to considering the authenticity of a generated trajectory and whether it is consistent with the historical trajectory, additional map information is introduced to establish a matching model between the generated trajectory and the current map. Experiments based on the Argoverse and nuScenes dataset are subsequently performed. The experimental results show that our prediction method outperforms state-of-the-art prediction systems, namely, TNT, PRIME and P2T. The strong coupling of the HD map significantly improves the reasonableness of the predicted trajectory.