Learning-based trajectory prediction models are increasingly being used in a wide range of AI applications, such as autonomous driving. However, existing methods usually ignore the potential distribution shift between the train and test environments. This inevitably results in an increased prediction error in the new domain. Towards this end, we present a novel model-agnostic student-teacher model that leverages the recent advances in self-training and utilizes predicted pseudo trajectories from the target domain in order to improve its generalization capabilities. More specifically, we propose to train the model using both trajectories from the source domain and predicted pseudo trajectories from the target domain. Since the predicted trajectories can be noisy, we weigh them by the epistemic uncertainty of the model using MC-dropout, giving more weight to the more certain ones. Additionally, we show that the domain gap can be reduced further by augmenting the source data. Experiments on the ETH and UCY datasets show the effectiveness of our framework on domain adaptation for pedestrian trajectory prediction.
Uncertainty-Aware Pseudo Labels for Domain Adaptation in Pedestrian Trajectory Prediction
24.09.2023
1323618 byte
Aufsatz (Konferenz)
Elektronische Ressource
Englisch
A context-aware pedestrian trajectory prediction framework for automated vehicles
ArXiv | 2021
|Context-aware Multi-task Learning for Pedestrian Intent and Trajectory Prediction
ArXiv | 2024
|