Data driven evaluation of crowds

Abstract

There are various techniques for simulating crowds, however, in most cases the quality of the simulation is measured by examining its "look-and- feel". Even if the aggregate movement of the crowd looks natural from afar, the behaviors of individuals might look odd when examined more closely. In this paper we present a data-driven approach for evaluating the behaviors of individuals within a simulated crowd. Each decision of an individual agent is expressed as a state-action pair, which stores a representation of the characteristics being evaluated and the factors that influence it. Based on video footage of a real crowd, a database of state-action examples is generated. Using a similarity measure, the queries are matched with the database of examples. The degree of similarity can be interpreted as the level of "naturalness" of the behavior. Essentially, this sort of evaluation offers an objective answer to the question of how similar are the simulated behaviors compared to real ones. By changing the input video we can change the context of evaluation. © Springer-Verlag 2009.