A Goal-oriented Trajectory Prediction Network

Predicting multimodal future trajectory of autonomous vehicles is crucial to making driving decisions, but it is challenging due to intention uncertainty and motion stochasticity. Existing models generally tackle this by predicting future behavior of agents based on extracted features and dense goal candidates which represent possible destinations. However, the quality of candidates limits prediction accuracy and densely sampling is not efficient enough. This paper proposes GoPNet, namely Goal-oriented Prediction Network which models latent proposals as goal representation for trajectory prediction. Specially, anchor-free and mode-specific proposals are utilized as the latent goal embedding rather than concrete goal coordinates. Compared with anchor-based methods, our model can not only reduce computation burden, but also preserve inherent diversity. In addition, proposals are interacted with scene context to generate middle proposal content features, which is supervised by an intermediate goal loss to learn more features about trajectory endpoints. Experiments on the large-scale driving dataset, Argoverse, show that GoPNet achieves better performance than other goal-conditioned methods. Besides, ablation study is conducted to validate the effectiveness of our model.