The idea of "safety in numbers" clearly has its limits. Each year, crowd-related disasters kill hundreds of people, and have typically been hard to prevent. But now an intervention may be at hand, thanks to crowd simulations developed by Paul M. Torrens, a geographer at Arizona State University. Torrens's computer simulations let planners drop a few thousand virtual people into riot scenes and burning buildings, then sit back and take notes. The specific scenarios Torrens creates could show firefighters how to save the most people, tell architects where to place exits or barriers in stadiums, and guide police forces in corralling unruly mobs.

While most traditional crowd simulations treat individuals as purely physical, with no social or emotional reactions, Torrens's model turns each individual into an "avatar" with an artificial mind. Avatars can plan their own route, adjust their path on the fly, and even respond to the body language of fellow cybercitizens who may be jostling them.

To provide his avatars with realistic-looking gait and body language, Torrens equips actors with markers on their bodies at key vertices and endpoints. As the person acts out different scenarios--walking down the street, taking part in an angry mob, running into a wall--a bank of high-res cameras records the position of the markers (the red circles in the diagram), a technique called motion capture. The computer program interpolates the connections ("edges") between the markers based on human skeletal bones, giving us the movement of the skeletal "rig" through space and time. Once digitized, the avatar's base behavioral data can be manipulated to change its size, speed, and strength.

With ray tracing, the skeletal rig can be rendered as a simple small-polygon figure or in a more complex human form. The rigs can then be wrapped in a textured "envelope" and rendered with lighting and shadows to make the scene look more realistic. The avatars are then infused with individual variables (behavioral algorithms for age, sex, mental health, and emotional state) that govern their movements and unique "personalities"--even though they all end up looking like agents from The Matrix.

And "agents" happens to be what Torrens rightfully calls them. He drops 300 of these agents on an urban street, where a convertible catches fire on the side of the road. The fire is generated by a physically realistic model, with smoke, embers, and mobile toxins subject to small eddies and turbulence between downtown buildings. The virtual pedestrians are subject to any number of real-world hazards--they can trip and fall and, if an ember lands on their hair or clothes, they can even catch fire. All are instructed to evacuate the area to safety on the right side of the image.

This suited throng of synthetic agents obediently rushes away from the burning car. Each agent is endowed with a sophisticated set of typical unconscious behaviors, but each is also programmed with a "personality" to override their instincts, demonstrating complex, rational behaviors that are derived from sociological theories of human behavior (e.g. pedestrians tend to drift right to avoid collisions--even in countries where they drive on the left). Constantly scanning his surroundings, an agent reacts to collisions, his neighbor's actions, and the ever-shifting dynamic of the crowd. Because of the heterogeneity of personalities and unique surroundings for each agent, the behavior of the group as a whole is unpredictable--just like real crowds under duress.

The individual nature of the agents makes the scene much more lifelike than the sims used in multimillion dollar-budget films like Lord of the Rings and King Kong. Crowd scenes in these movies use animation algorithms based on flocking activity, in which each character moves in a preprogrammed way depending upon how her neighbors move.

In two scenarios, the crowd is instructed to either run (left) or walk calmly (right) to a single, narrow exit. When navigating, the area scanned by each agent changes size and shape based on the speed and density of the crowd. When the runners encounter traffic, they focus on nearby collisions, ignoring far-off features, and individual behavior yields to crowd dynamics. The agents interpret subtle signals in their neighbors' body language and abruptly adjust to avoid smashing into people and other obstacles best avoided. The small scanning area and constant adjustments in the running scenario cause spontaneous jams that ripple backwards against the tide of crowd flow. The walkers have a greater buffer of personal space and are able to proceed at a more constant pace, allowing for a larger scanning area.

As the crowd reaches the exit in the running scenario, they form a large wedge-like mass around the exit, restricting evacuation to all but a trickle. Like a mosh pit at a Pantera concert, the frenzied crowd compresses and expands, resulting in high pressure areas and serious injuries. In the walking scenario, the area still gets congested but the gridlocked crowd directly in front of the exit is less dense; the herd is more like the head-bobbing fans at the back of the auditorium, exerting less pressure on each other and causing fewer injuries.

One design that can expedite the flow of crowds through constrained corridors is, ironically, a column placed slightly off-center in front of the exit. This allows for "bubbles" to form in the crowd, dampening congestion ripples in much the same way that rush-hour traffic signals on entrance ramps smooth vehicular traffic flow. Here, the model is shown in "empirical mode" with minimal graphics. One agent is selected (yellow) and his speed and injury status are displayed as he navigates through the crowd. The circles show the spatial footprint of the agents and the triangles show their orientation.

Probes in the simulation report data every 60th of a second--the reaction time for human movement--to allow Torrens to track every action and interaction. Each colored line represents the path of an individual agent. By subjecting the space-time signature of every individual to spatial and social network analyses, Torrens can identify the sources of gridlock in both the crowd and the urban infrastructure. Torrens also considers the toruosity--the crookedness--of each agent's path. The straighter the path is from the start point to the goal, the more efficient the movement; if a path deviates from "normal" behavior, the agent has panicked and is moving inefficiently.

Read the data article as it appears in the April, 2008 Issue.