Geometric collision avoidance for heterogeneous crowd simulation

Simulation of human crowds can create plausible human trajectories, predict likely flows of pedestrians, and has application in areas such as games, movies, safety planning, and virtual environments. This dissertation presents new crowd simulation methods based on geometric techniques. I will show how geometric optimization techniques can be used to efficiently compute collision-avoidance constraints, and use these constraints to generate human-like trajectories in simulated environments. This process of reacting to the nearby environment is known as local navigation and it forms the basis for many crowd simulation techniques, including those described in this dissertation. Given the importance of local navigation computations, I devote much of this dissertation to the derivation, analysis, and implementation of new local navigation techniques. I discuss how to efficiently exploit parallelization features available on modern processors, and show how an efficient parallel implementation allows simulations of hundreds of thousands of agents in real time on many-core processors and tens of thousands of agents on multi-core CPUs. I analyze the macroscopic flows which arise from these geometric collision avoidance techniques and compare them to flows seen in real human crowds, both qualitatively (in terms of flow patterns) and quantitatively (in terms of flow rates). Building on the basis of these strong local navigation models, I further develop many important extensions to the simulation framework. Firstly, I develop a model for global navigation which allows for more complex scenarios by accounting for long-term planning around large obstacles or emergent congestion. Secondly, I demonstrate methods for using data-driven approaches to improve crowd simulations. These include using real-world data to automatically tune parameters, and using perceptual user study data to introduce behavioral variation. Finally, looking beyond geometric avoidance based crowd simulation methods, I discuss methods for objectively evaluating different crowd simulation strategies using statistical measures. Specifically, I focus on the problem of quantifying how closely a simulation approach matches real-world data. I propose a similarity metric that can be applied to a wide variety of simulation approaches and datasets. Taken together, the methods presented in this dissertation enable simulations of large, complex humans crowds with a level of realism and efficiency not previously possible.