Intention-aware Transformer with Adaptive Social and Temporal Learning for Vehicle Trajectory Prediction

Accurate prediction of the trajectory of surrounding vehicles is crucial to autonomous driving for path planning and collision avoidance. In this paper, we propose a novel transformer-based model with adaptive social and temporal learning for trajectory prediction. In order to model social interaction between vehicles at each historical timestamp, an enhanced graph attention feature aggregation mechanism combing hidden feature and explicit relative spatial relation is developed. Further, the social and temporal dependency across different timestamps is captured by multi-head self-attention with an extra learnable "intention token". To achieve multi-modal trajectory prediction, we implement intention-aware transformer decoder of driving behavior, the intention recognition and trajectory. Experiments on large-scale benchmark datasets verify that our model achieves better performance comparing with some state-of-the-art trajectory prediction models.