Trending Paths: A New Semantic-Level Metric for Comparing Simulated and Real Crowd Data

Area of Science:

• Computer Vision
• Artificial Intelligence
• Crowd Simulation

Background:

• Evaluating crowd simulation fidelity is crucial for real-world applicability.
• Existing metrics focus on low-level (e.g., trajectories) or global (e.g., densities) features.
• A need exists for a semantic-level evaluation approach.

Purpose of the Study:

• To propose the first semantic-level crowd evaluation metric.
• To introduce a method for analyzing and comparing real and simulated crowd data based on latent patterns.
• To develop a robust and versatile metric for crowd behavior analysis.

Main Methods:

• Utilizing unsupervised clustering via non-parametric Bayesian inference to learn latent Path Patterns.
• Introducing a new Stochastic Variational Dual Hierarchical Dirichlet Process (SV-DHDP) model.
• Computing pattern fidelity against a reference for comparative analysis.

Main Results:

• The proposed method provides a rich visualization of crowd behavior through learned patterns.
• The metric allows for comparison between different crowd simulation algorithms and real data.
• Demonstrated robustness to noise and fewer assumptions compared to existing metrics.

Conclusions:

• The novel semantic-level metric offers a valuable alternative for crowd data comparison.
• The approach is applicable to diverse data types and improves upon existing evaluation methods.
• This work advances the field of crowd simulation evaluation through semantic pattern analysis.