Scalable pedestrian simulation for virtual cities

Abstract

Most of the common approaches for the pedestrian simulation, used in the Graphics/VR community, are bottom-up. The avatars are individually simulated in the space and the overall behavior emerges from their interactions. This can lead to interesting results but it does not scale and can not be applied to populating a whole city. In this paper we present a novel method that can scale to a scene of almost any size. We use a top-down approach where the movement of the pedestrians is computed at a higher level, taking a global view of the model, allowing the flux and densities to be maintained at very little cost at the city level. This information is used for stochastically guiding a more detailed and realistic low level simulation when the user zooms in to a specific region, thus maintaining the consistency. At the heart of the system is an iterative method that models the flow of avatars as a random walk. People are moved around a graph of nodes until the model reaches a steady state which provides feedback for the avatar low level navigation at run time. The Negative Binomial distribution function is used to model the number of people leaving each node while the selected direction is based on the popularity of the nodes through their preference-factor. The preference-factor is a function of a number of parameters including the visibility of a node, the events taking place in it and so on. An important feature of the low-level dynamics is that a user can interactively specify a number of intuitive variables that can predictably modify the collective behavior of the avatars in a region

the density, the flux and the number of people can be selectively modified. Copyright 2004 ACM.