Authors: Guipeng Xin, Duanfeng Chu*, Liping Lu, Xigang Wu, and Peichang Ye
Trajectory prediction plays a crucial role in autonomous driving, aiming to forecast the future reasonable motion trajectories of traffic participants. Most existing methods model agents uniformly, neglecting the behavioral heterogeneity among different agents. In this paper, we propose a novel Difficulty-Guided Feature Enhanced Network (DGFNet) to leverage the predictive difficulty differences among agents for trajectory prediction. Firstly, we employ spatio-temporal feature encoders to capture rich temporal and spatial features. Secondly, we utilize difficulty-guided feature enhancement to obtain reliable future trajectories in the initial predictions, providing future trajectory features for global fusion. Thirdly, the fused features are further input into the final predictor to generate predicted trajectory distributions for multiple participants. Experimental results demonstrate that our DGFNet achieves advanced performance on the Argoverse 1 motion prediction benchmark. Our implementation with additional visualizations is available at https://github.com/XinGP/DGFNet.
Code will be released soon.
This repository is licensed under Apache 2.0.