TranSense: A transformer-based trajectory prediction method for multi-layer lane intent estimation

Accurately predicting the future trajectories of other traffic participants represents a critical capability for autonomous vehicles. Current methods lack a robust perception model capable of handling complex scenarios. To this end, this paper introduces Transense, which employs the Transformer architecture to analyze agents’ lane selection intentions. Through iterative refinement, Transense effectively identifies pivotal factors influencing agents’ lane choices. Furthermore, a segmented refinement approach is applied in the decoder phase, pre-decoding guiding trajectories and innovatively integrating an attention mechanism to facilitate interaction between the guiding trajectories and scene information, thereby enabling trajectory correction. This methodology preserves temporal information while enhancing prediction accuracy. Experimental validation on the nuScenes and Argoverse datasets demonstrates that our method outperforms existing approaches.