Leveraging Future Relationship Reasoning for Vehicle Trajectory Prediction
– Published Date : TBD
– Category : Trajectory prediction
– Place of publication : International Conference on Learning Representations (ICLR) 2023
Understanding the interaction between multiple agents is crucial for realistic and plausible vehicle trajectory prediction. Accordingly, existing methods tried to model and predict the interaction using observed past trajectories of agents with pooling, attention, or graph-based methods. However, we observed that they easily fail under complex road structures. It is because they do not explicitly utilize the map information for predicting the relationship, and they only model the relationship between vehicles in a deterministic manner, not in a stochastic manner. In this paper, we propose a new method to model a stochastic future relationship among agents using the lane information contained in the map. Our method first predicts a probability of lane-level waypoint occupancy of vehicles. Then the proposed method utilizes the temporal probability of passing the adjacent lanes to learn the interaction between agents. In addition, we model the interaction employing probabilistic distribution. This distribution is learned by the posterior distribution of interaction from GT future trajectory. As a result, we validate our method on popular trajectory prediction datasets: nuScenes and Argoverse. The code will be available in public upon acceptance.