MIFT: Multi-Agent Motion Prediction with the Interaction of Future Trajectories

Motion prediction is crucial for enabling autonomous driving systems to intelligently navigate complex traffic scenarios and challenges, thereby enhancing safety, efficiency, comfort, and adaptability. From the driver's or pedestrian's own point of view, its true motion traj ectory is not fixed, but is adjusted or changed in real time according to the motion trajectory of the surrounding agent. Therefore, for multi-agent motion prediction, considering the possible interaction of future traj ectories between them and modifying traj ectories is more in line with the motion of agents in real scenarios. In this paper, we propose MIFT: on the basis of the conventional framework of encoder-interaction-decoder, we add a future trajectory interaction module to output the correction of future predicted traj ectories. The experimental results of argoverse2 motion prediction dataset show that MIFT framework has achieved good performance on highly competitive motion prediction benchmarks, and the rationality of the existence of future trajectory interaction module is verified by ablation experiments.