We develop models for simulating realistic, believable crowds. To this end, we identify and address fundamental limitations in how individuals in a crowd are represented and controlled. Our research results have widespread application in visual effects, games, urban planning, architecture design, as well as disaster and security simulation. Our research deliverables are provided as part of SteerSuite, an open-source platform for simulating and evaluating crowd behavior.
Data-Driven Modeling of Human Crowds
In order to use computational models of crowd movement for evaluating and informing building designs, it is imperative that these models capture the essence of real human crowd interactions, while generalizing to novel environment layouts. Current approaches make simplifying assumptions that prevent the models from accounting for the locomotion affordances and constraints exhibited by humans in high-density interactions. Additionally, these models are manually tuned to replicate specific crowd phenomena with no existing mechanisms for automatically fitting these models to capture the essence of real crowd datasets in a representative and general fashion. We leverage our research in autonomous virtual humans to serve as a basis for computational models of human crowd behavior. Our proposed models facilitate the extrapolation of the crowd dynamics in unseen situations (such as those featuring increased density or new environment layouts) while preserving the crowd characteristics observed in reality.
Statistical Analysis of Crowd Dynamics
Our ability to interpret emergent patterns in crowd situations is still limited. A key question is how does the environment shape the collective patterns that emerge in aggregate crowd dynamics. We have introduced a variety of measures and benchmarks to quantitatively analyze, evaluate, and compare both real and synthetic crowd data, with application in behavior and anomaly detection. Our solutions for modeling and analysing crowds have been demonstrated for predictive crowd analytics, crowd behavior optimization, and environment analysis. We are developing a novel information-theoretic framework to systematically extract arbitrarily complex statistical dependencies from raw crowd motion patterns and the environment configuration. To mitigate the computation complexity of expensive simulations, we are exploring deep networks to automatically predict the relation between the environment and crowd configuration, and these information measures.
Crowd Behavior Optimization
We have developed a framework for automatically optimizing the parameters of a crowd simulation algorithm to meet any user-defined criteria. Our tools can be used to predictively evaluate the optimal crowd behavior in evacuations, and other stressful situations, and run “what-if” scenarios for crowd behavior management.