We present a technique for modeling large-scale urban environments for virtual human simulation. Our approach, which involves a set of hierarchical data structures, supports the efficient interaction between numerous autonomous pedestrians and their environment, including perceptual processing and path planning for the purposes of goal directed navigation. As a specific implementation of our approach, we develop an environmental model of a large train station and demonstrate its ability to support the real-time simulation of more than a thousand autonomous pedestrians engaged in a broad variety of individual and group behaviors appropriate to their large-scale urban environment.