Diverse and plausible vehicle trajectory prediction using two-phase learning

Abstract

Pedestrian and vehicle path prediction is emerging as one of the essential algorithms in the field of robotics and autonomous driving. In order to implement safe autonomous driving in consideration of future uncertainties, diverse and plausible trajectories must be predicted. However, the most existing trajectory prediction works focus on predicting the answers in datasets. In addition, most of them try to learn the diversity and validity which have trade-off between them at once so they miss both the diversity and validity. To overcome the trade-off between them, this thesis propose two-phase learning to make model learn diversity and plausibility separately. The proposed framework outperforms other methods for diversity and plausibility while maintaining the comparable validity with other methods in Argoverse and nuScenes dataset.