Roland Geraerts

Utrecht University, Netherlands

Title: Crowd simulation - A computational model of human navigation



A huge challenge is to simulate tens of thousands of virtual characters in real-time where they pro-actively and realistically avoid collisions with each other and with obstacles present in their environment. This environment contains semantic information (e.g. roads and bicycle lanes, dangerous and pleasant areas), is three-dimensional (e.g. contains bridges where people can walk over and under as well) and can dynamically change (e.g. a bridge partially collapses or some fences are removed). We show how to create a generic framework centered around a multi-layered navigation mesh and how it can be updated dynamically and efficiently for such environments. Next, we show how (groups of) people move, avoid collisions and coordinate their movements, based on character profiles and semantics. We run our simulations in realistic environments (e.g. soccer stadiums or train stations) and game levels to study the effectiveness of our methods. Finally, we demonstrate our software package that integrates this research. Why would we need to simulate a crowd? The results can be used to decide whether crowd pressures do not build up too much during a festival such as the Lode Parade; to find out how to improve crowd flow in a train station; to plan escape routes for use during a fire evacuation; to train emergency personnel to deal with evacuation scenarios; to study a range of scenarios during an event; or to populate a game environment with realistic characters. After this presentation you'll understand why state-of-the-art crowd simulations need a more generic and efficient representation of the navigable areas, why speed and extendability is obtained by splitting up the simulation in at least five different levels, why we need a paradigm shift from graph-based to surface-based navigation, and why a path planning algorithm should NOT compute a path.