Abstract:

In vehicle trajectory prediction, it is important to consider long-term and short-term intentions. Long-term intention refers to the final destination a vehicle is trying to reach within a predetermined time frame, while short-term intention refers to the immediate decisions made by the vehicle as it moves toward its goal. These decisions can be influenced by interactions with other vehicles. In this paper, we propose the Dual-Sampling Strategy (DSS) that takes into account both types of intentions probabilistically. Since it is reasonable to assume that a vehicle moves along lanes, we use lanes as strong prior knowledge to model the two types of intentions. Inspired by the fact that experienced drivers focus on other vehicles that could potentially occupy same lanes with themselves, we also propose the Future Relation Module (FRM), which considers lane-level proximity to model the relationships between vehicles. With the proposed sampling methods, we achieve state-of-the-art performance with a large margin for all metrics on the online benchmark dataset.