Dual-Alignment Domain Adaptation for Pedestrian Trajectory Prediction

Abstract

Predicting the plausible future paths of pedestrians is essential for human-involved applications (e.g., autonomous driving and service robotics). Existing pedestrian trajectory prediction methods mainly focus on the performance of multi-scene trained models in single-scene tests, neglecting the cross-scene knowledge differences in practice. To address this issue, we propose a generic dual-alignment framework for pedestrian trajectory prediction. Concretely, we analyze the domain difference at macro and micro scales and mitigate them respectively: at macro scale, an attention-based temporal convolutional generative model transfers the paths of pedestrians and their interaction information from the source domain to the target domain to align the data-level distributions; at micro scale, an auxiliary adversarial network is integrated to assist in training the prediction network to align the feature-level domain-invariant knowledge. Cross-domain experiments demonstrate that our approach significantly improves the performance of existing pedestrian trajectory prediction benchmarks (up to 53.5%) and outperforms previous domain adaptive works (up to 41.7%).