Multi-Perception Crowd: Learning to combine entity and implicit perception for diverse crowd simulation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Simulation

Background:

• Crowd simulation requires balancing collision avoidance with navigation preferences.
• Existing methods often struggle with complex environments and intuitive authoring.

Purpose of the Study:

• To develop a reinforcement learning framework for crowd simulation.
• To enable intuitive authoring of crowd behaviors through interactive tools.
• To achieve robust generalization across complex scenes.

Main Methods:

• Integrating entity perception (hard constraints) and implicit environmental perception (soft constraints via suitability maps).
• Employing a modular, two-phase training strategy with curriculum-based environmental templates.
• Developing an intuitive crowd authoring tool for real-time behavior modification.

Main Results:

• The framework successfully balances collision avoidance and navigation preferences.
• Generated crowd behaviors align with real-world pedestrian traffic patterns.
• Emergent social behaviors, like desire path formation, were observed.

Conclusions:

• The proposed framework offers a robust, learning-based approach to crowd simulation.
• It supports heterogeneous navigation modeling and intuitive authoring for interactive applications.
• The method enhances realism and control in crowd simulation.