Context-aware motion diversification for crowd simulation.

Abstract

Traditional crowd simulation models typically focus on navigational pathfinding and local collision avoidance. Little research has explored how to optimally control individual agents' detailed motions throughout a crowd. A proposed approach dynamically controls agents' motion styles to increase a crowd's motion variety. The central idea is to maximize both the style variety of local neighbors and global style utilization while maintaining a consistent style for each agent that's as natural as possible. To assist runtime diversity control, an offline preprocessing algorithm extracts primitive motions from a motion capture database and stylizes them. This approach can complement most high-level crowd models to increase realistic variety. Four experiment scenarios and a user evaluation demonstrate this approach's superior flexibility compared to traditional random distribution of motion styles. The Web extra is a video demonstrating a military-march simulation.