Dynamic-group-aware networks for multi-agent trajectory prediction with relational reasoning

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Precise trajectory prediction requires understanding complex agent interactions.
• Existing methods often overlook dynamic, group-based relationships, limiting predictive accuracy.

Purpose of the Study:

• To introduce DynGroupNet, a novel network for modeling time-varying interactions in dynamic scenes.
• To enhance trajectory prediction by capturing both pairwise and group-wise agent relationships.
• To develop a prediction system that forecasts socially plausible trajectories using dynamic relational reasoning.

Main Methods:

• DynGroupNet models time-varying, pairwise, and group-wise interactions without direct supervision.
• A prediction system utilizes Gaussian mixture models, multiple sampling, and refinement for diverse, stable, and smooth trajectory forecasting.
• Dynamic relational reasoning is employed to infer interaction strength and category.

Main Results:

• DynGroupNet successfully captures time-varying group behaviors and infers interaction dynamics.
• The system significantly outperforms state-of-the-art methods, showing improvements of 28.0% (NBA), 34.9% (NFL), and 13.0% (SDD) in Final Displacement Error (FDE).

Conclusions:

• DynGroupNet provides a robust framework for understanding and predicting complex agent interactions in dynamic environments.
• The proposed method enhances prediction accuracy and social plausibility, setting a new benchmark in trajectory prediction research.