Pedestrian Motion Characteristic-based Trajectory Prediction from the Driver’s Perspective

In traffic scenarios, pedestrians navigate to their destinations while avoiding obstacles. Autonomous vehicles must predict pedestrian future positions and adjust their paths to prevent accidents. Previous trajectory prediction methods commonly use a static camera's view, not directly suitable for autonomous driving. This study proposes a method from a vehicle's perspective to predict pedestrian trajectories. Vehicle motion information is fused with pedestrian features, then used as input for subsequent prediction module processing. Two modules, MFM and SIM, are proposed to solve the interaction problem between individual pedestrian features and neighborhood pedestrians respectively. Our method's performance is verified using the MOT16 multi-person dataset and the Daimler pedestrian path prediction benchmark dataset. Compared to the existing FPL method, our approach reduces the mean displacement error (ADE) by 8% and the final displacement error (FDE) by 10%. It provides a crucial step towards safe coexistence of autonomous vehicles and pedestrians, by offering both theoretical and tested means for predicting the future location of pedestrians.