GCSFNet: A self-supervised fusion network for trajectory prediction in complex traffic scenarios

With the rapid development of intelligent transportation and autonomous driving technologies, vehicle trajectory prediction, as a key task for ensuring traffic safety and efficiency, has become a crucial link in improving the decision-making performance of autonomous vehicles. To address the issue that traditional supervised learning is highly dependent on large amounts of labeled data, this paper proposes a vehicle trajectory prediction method based on a self-supervised learning mechanism, aiming to more fully explore the feature information of unlabeled data and enhance the model’s ability to fuse information in complex scenarios. The method is designed with a Global Channel-Spatial Fusion (GCSF) module to improve the extraction capability of multi-source feature information; it conducts multi-dimensional modeling by effectively integrating historical trajectory information, lane geometry information, and position information, enriches the representation ability of feature information, and achieves high-precision trajectory prediction. In the pre-training phase, it further incorporates Spatial Information (SI) enhancement to construct a multi-level self-supervised pre-training framework, realizing the feature enhancement of historical and future trajectory information for the pre-trained model and the establishment of a deep-level model. Experiments based on the Argoverse 2 dataset demonstrate that the proposed method outperforms existing mainstream methods in multiple prediction metrics; in particular, the minADE 1 and minFDE 1 metrics are improved by 1.3% and 3.0% respectively compared with the baseline models, which effectively enhances the generalization ability and robustness of the model and has important reference value for the intelligent decision-making of future autonomous driving systems.