A Clustering Based Approach for Realistic and Efficient Data-Driven Crowd Simulation

TL;DR

This paper introduces a hierarchical, data-driven crowd simulation method that improves realism and efficiency by classifying agents into interaction patterns and selecting actions accordingly, outperforming existing models.

Contribution

It presents a novel hierarchical approach combining rule-based and data-driven models for more realistic and efficient crowd simulation, with automatic pattern discovery and behavior generation.

Findings

Outperforms state-of-the-art models in prediction accuracy.

Operates at interactive real-time speeds.

Handles various crowd densities and directions effectively.

Abstract

In this paper, we present a data-driven approach to generate realistic steering behaviors for virtual crowds in crowd simulation. We take advantage of both rule-based models and data-driven models by applying the interaction patterns discovered from crowd videos. Unlike existing example-based models in which current states are matched to states extracted from crowd videos directly, our approach adopts a hierarchical mechanism to generate the steering behaviors of agents. First, each agent is classified into one of the interaction patterns that are automatically discovered from crowd video before simulation. Then the most matched action is selected from the associated interaction pattern to generate the steering behaviors of the agent. By doing so, agents can avoid performing a simple state matching as in the traditional example-based approaches, and can perform a wider variety of steering behaviors as well as mimic the cognitive process of pedestrians. Simulation results on scenarios with different crowd densities and main motion directions demonstrate that our approach performs better than two state-of-the-art simulation models, in terms of prediction accuracy. Besides, our approach is efficient enough to run at interactive rates in real time simulation.