Parallel implementation of large scale crowd simulation 


When designing entrances and exits for a stadium, a transit station, ships or similar structures, it is
important to optimize the flow of people which in turn requires that flow to be simulated. In addition to
that, the police, the military, firefighters and others who have to deal with various types of crowds either
to contain and control them or to evacuate them, are more and more using simulated training in their
education and they also have a need for virtual crowds.
Various tools exist to perform crowd simulation, but they are generally not real-time and can be quite
time consuming for a large scale simulation.  The national football stadium, Parken, has room for over
40.000 persons so a full scale simulation dealing with an evacuation scenario will be computationally
expensive.  When used for interactive virtual training, real-time is an actual requirement and not just a
convenience.
In order to make large scale simulation possible in real-time, or almost real-time, we will implement
a model for human crowd behavior on a parallel processing platform using CUDA (a c based API for
parallel processing) and evaluate the accuracy of the model along with its ability to simulating full scale
scenarios with very large crowds.
Accuracy will be evaluated by comparison with various text book crowd behaviors such as the forma-
tion of ”lanes” and ”vortices” when two or more crowds moving in different directions meet.
We contribute with a clear relation between fluid simulation and human crowds where the controlling
forces inside a crowd are connected with their physical fluid counterparts.  The implementation results
in a a method with a time complexity that depends on the size of the domain and not on the number of
agents.

The paper can be downloaded here
Ċ
Thomas Grønneløv,
Apr 12, 2011, 12:31 PM