MDST-DGCN: A Multilevel Dynamic Spatiotemporal Directed Graph Convolutional Network for Pedestrian Trajectory...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Pedestrian trajectory prediction is crucial for applications like autonomous driving and urban planning.
• Accurate prediction relies heavily on understanding complex social interactions between pedestrians.
• Existing models often struggle to capture the multifaceted nature of these interactions.

Purpose of the Study:

• To develop a novel deep learning model for comprehensive pedestrian trajectory prediction.
• To effectively model multilevel social interactions among pedestrians.
• To improve prediction accuracy in diverse crowd densities and time scales.

Main Methods:

• Proposed a multilevel dynamic spatiotemporal directed graph convolutional network (MDST-DGCN).
• Incorporated a motion encoder for individual pedestrian features.
• Utilized a multilevel dynamic spatiotemporal directed graph encoder (MDST-DGEN) for adaptive fusion of social interactions.
• Employed a motion decoder for generating future trajectories.

Main Results:

• The MDST-DGCN model achieved state-of-the-art performance on public datasets.
• Demonstrated superior accuracy in both long-term and short-term trajectory predictions.
• Showcased effectiveness in both high-density and low-density crowd scenarios.

Conclusions:

• The proposed MDST-DGCN effectively captures multilevel social interactions for improved pedestrian trajectory prediction.
• The model offers a significant advancement in predicting pedestrian movement in complex environments.
• This approach holds promise for enhancing safety and efficiency in pedestrian-heavy areas.