MTP-STG: Spatio-Temporal Graph Transformer Networks for Multiple Future Trajectory Prediction in Crowds

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Predicting multiple future pedestrian trajectories is vital for autonomous driving and robotic motion planning.
• Current methods often neglect distant environmental influences and struggle with trajectory smoothness and temporal consistency.
• Understanding complex crowd dynamics requires advanced modeling techniques.

Purpose of the Study:

• To propose a multimodal trajectory prediction model for crowd scenarios.
• To enhance trajectory prediction by incorporating distant spatial information and ensuring temporal consistency.
• To achieve state-of-the-art performance in predicting multiple future pedestrian paths.

Main Methods:

• Utilized spatio-temporal graphical attention networks and multimodal data augmentation.
• Generated simulated pedestrian trajectory data using CARLA for multiview analysis.
• Employed a multitarget detection and tracking algorithm and a Memory Storage Module for trajectory smoothing.

Main Results:

• The proposed MTP-STG model demonstrated superior performance in predicting multiple future trajectories.
• Effectively modeled scaled spatial interactions among pedestrians in crowded settings.
• Achieved state-of-the-art results compared to existing trajectory prediction methods.

Conclusions:

• The developed multimodal trajectory prediction model significantly advances the field.
• Integrating spatio-temporal graph attention networks enhances prediction accuracy and consistency.
• The approach offers a robust solution for complex crowd navigation scenarios.