Who Should Have Been Focused: Transferring Attention-Based Knowledge from Future Observations for Trajectory Prediction
Abstract
To ensure safe driving of autonomous vehicles, we improve the prediction of surrounding vehicles' trajectories. In particular, to solve the problem that existing prediction models have difficulty resolving intentional uncertainties by relying on past and present observations, this study proposes the following approaches:
Teacher-Student Learning: The teacher model uses the future trajectory information of other agents to identify other agents that the target agent should focus on, and transfers this attentional knowledge to the student model so that the student model can learn to perform more accurate predictions with only past and present information.
Lane-guided Attention Module, LAM: It improves the accuracy of trajectory prediction by reducing the uncertainty of the trajectory by utilizing the lane information near the predicted trajectory and improving the agent-map interaction.
- ConferenceICPR
- TopicTrajectory Prediction
- Year Published2024
Abstract
To ensure safe driving of autonomous vehicles, we improve the prediction of surrounding vehicles' trajectories. In particular, to solve the problem that existing prediction models have difficulty resolving intentional uncertainties by relying on past and present observations, this study proposes the following approaches:
Teacher-Student Learning: The teacher model uses the future trajectory information of other agents to identify other agents that the target agent should focus on, and transfers this attentional knowledge to the student model so that the student model can learn to perform more accurate predictions with only past and present information.
Lane-guided Attention Module, LAM: It improves the accuracy of trajectory prediction by reducing the uncertainty of the trajectory by utilizing the lane information near the predicted trajectory and improving the agent-map interaction.