Abstract:

To address limitation in prediction accuracy and data utilization efficiency of supervised learning-based trajectory prediction models， a trajectory prediction model and a general self-supervised pretraining strategy are proposed. Firstly， a lightweight trajectory prediction model based on Transformer is established to extract temporal-spatial features while modeling interaction relationship. Secondly， three types of masks， namely motion information temporal mask， road information spatial mask， and interaction relationship mask， are designed for self-supervised pre-training tasks on the model to enhance the model's ability to extract general scene features. Finally， pretraining weights are used as initialization parameters for supervised learning fine-tuning in downstream tasks. Experimental results on the Argoverse2 Motion Forecasting dataset show that the model can effectively reconstruct traffic scenes in pretraining tasks. The introduction of self-supervised pretraining improves prediction accuracy and data utilization efficiency. Moreover， it exhibits universality for different prediction tasks， achieving a 3.3% and 3.7% improvement in the minFDE₆ for single-agent and multi-agent trajectory prediction tasks， respectively.

Key words: autonomous driving, trajectory prediction, self-supervised pretraining