MITNet: Multi-intention Trajectory Network for Motion Prediction

Predicting the future behavior of mobile agents is critical for real-world applications. It is challenging because the agent's intentions and corresponding behaviors are unknown and multimodal in nature. Recent data-driven motion prediction methods attempt to learn to regress the exact future location or its distribution directly from large amounts of trajectory data. However, these methods are difficult to provide multimodal prediction and synthesis of physical constraints such as traffic rules and movable areas. In this work we propose a novel three-stage motion prediction framework, Multi-intention Trajectory Network (MITNet). Experiments on our logistic dataset and large-scale public trajectory prediction datasets show that MITNet achieves state-of-the-art results both quantitatively and qualitatively.