Dense Crowds of Virtual Humans

This thesis presents a novel crowd simulation method `Torso Crowds', aimed at the simulation of dense crowds. The method is based on the results of user studies and a motion capture experiment, which are also described in this thesis. Torso Crowds introduces a capsule shape to represent people in the crowd, and supports different behaviour for actively manoeuvring people and people that have no incentive to move. Throughout history, people have moved from the country into cities. The number of people in the world keeps growing, resulting in the highest local population densities of all time. Furthermore, the scale of events and venues such as football stadiums and concert halls increases. As a result, we see more, larger crowds of people; their safety and well-being becomes more important, while at the same time becoming more difficult to predict. Recent history has shown how quickly a situation can take a turn for the worse when a large crowd is involved. Examples are the crushing disasters at the Hillsborough stadium in 1989 and the Love Parade in Duisburg in 2010. Computer simulations of such events can help in various ways. Police, riot control and medical staff can use real-time crowd simulation for training purposes, and test the effectiveness of their crowd management strategies in a safe and controlled virtual environment. City council members and event planners can also use simulations to investigate crowd flow and find potentially dangerous bottlenecks. Even before construction starts, building designs can be evaluated for crowd flow and evacuatability. In these simulations, it is vital that the behaviour of the simulated crowd is representative of that of real people. The main motivation for the research reported on in this thesis is the desire to increase realism in simulations of dense crowds. Most simulation methods use very simple, symmetrical shapes, such as discs or points, to represent members of the crowd. Their orientation is directly related to their velocity, and as a result characters rotate instantaneously. The common disc shape also results in non-humanoid behaviour in dense situations; when many agents are pressed into a small area, this shape becomes indirectly visible; their motion reminds of beer bottles pushing against each other in a factory. This led to the use of a capsule shape, which also allows for more realistic motion by allowing side-stepping, walking backward, and twisting the torso to manoeuvre through narrow openings in the crowd. Various metrics for comparing crowd simulation performance focus on the avoidance of collisions. Although such avoidance in itself is realistic, a simulation of a dense crowd that shows no collisions at all is not. This led us to produce a fast collision detection technique, based on hierarchies of cylinders. Single cylinders are often used to detect collisions between virtual characters; by refining the single cylinder using gradually smaller cylinders we detail this shape, and make the collision detection more precise