A Novel Graph-Based Trajectory Predictor With Pseudo-Oracle.

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Pedestrian trajectory prediction is crucial for autonomous systems like self-driving cars.
• Existing methods struggle to capture complex social interactions and obstacle avoidance behaviors.
• Future uncertainties in pedestrian movement pose significant prediction challenges.

Purpose of the Study:

• To develop an advanced pedestrian trajectory prediction model.
• To integrate obstacle avoidance experiences (OAEs) into prediction models.
• To enhance prediction accuracy in dynamic and socially interactive environments.

Main Methods:

• An encoder-decoder architecture using Long Short-Term Memory (LSTM) with temporal attention.
• A graph-based attention mechanism incorporating OAEs for modeling interactions.
• A novel pseudo-oracle predictor to generate an informative latent variable for handling future uncertainties.
• Multimodal output generation for diverse future trajectory possibilities.

Main Results:

• The Graph-based Trajectory Predictor with Pseudo-Oracle (GTPPO) achieved state-of-the-art performance on benchmark datasets (ETH, UCY, Stanford Drone).
• Qualitative evaluations demonstrated successful prediction of sudden pedestrian motion changes.
• The model effectively captures pedestrian motion patterns, social interactions, and future uncertainties.

Conclusions:

• GTPPO offers a significant advancement in pedestrian trajectory prediction accuracy and robustness.
• Integrating OAEs and a pseudo-oracle predictor enhances the model's ability to anticipate pedestrian behavior.
• The findings suggest GTPPO can effectively 'peek into the future' of pedestrian movements.