A data-driven stacking fusion approach for pedestrian trajectory prediction

This paper systematically investigates pedestrian trajectory prediction through a data-driven stacking fusion approach. Firstly, a novel Attention Mechanism-Long Short-Term Memory Network (Att-LSTM) is presented for pedestrian trajectory prediction, pedestrian heterogeneity and pedestrians-dynamic vehicles interactions are considered. Then, a Modified Social Force Model (MSFM) is developed for pedestrian trajectory prediction. The collision avoidance with conflicting dynamic vehicles and pedestrians, the influence of crosswalk boundary and pedestrian heterogeneity are considered. Finally, a data-driven stacking fusion model based on the Att-LSTM and MSFM is developed, and ridge model is used to prevent model overfitting and enhance model robustness. Moreover, traffic data of an un-signalised crosswalk is collected; the non-measurable parameters are calibrated through the Maximum-Likelihood Estimation. The model evaluation results show that the stacking fusion model performs better than the existing methods, which make it possible for autonomous vehicle to present great feasibility for improving pedestrian safety and traffic efficiency.

Automated summary

This paper systematically investigates pedestrian trajectory prediction through a data-driven stacking fusion approach. The proposed Att-LSTM and MSFM models outperform existing methods in terms of pedestrian trajectory prediction. The results show that the stacking fusion model has great feasibility for improving pedestrian safety and traffic efficiency in autonomous vehicles.