News
Computing Profession

Avoiding the Crush

Posted
Tracking the behavior of members of a crowd.
Tracking the behavior of members of a crowd.
Researchers have been developing models that mimic how people move in large groups.

Earthquakes, floods, and hurricanes take lives around the world, due to their unpredictable nature and massive power. Humans are responsible for similar levels of carnage through human crushes or stampedes; for example, more than 2,200 Muslim pilgrims were killed in Mina, Saudi Arabia, during an overcrowding incident during the Hajj pilgrimage in September 2015. On New Year’s Eve in 2014, 36 people died and 42 were injured after a stampede during celebrations at The Bund, in Shanghai, China. During the Hindu Navratri festival in October 2013, a stampede at the Ratangarh Mata Temple in India killed 115 people and injured more than 100.

To address such public safety issues, researchers have been developing computer models that mimic how people actually move when in large groups. Their ultimate goal: to develop reliable algorithms that can feed into models to accurately predict how crowds move, then design physical spaces to safely accommodate and control that movement to prevent deadly crushes.

Dinesh Manocha, a professor at the University of North Carolina at Chapel Hill, has been studying crowd behavior for more than eight years, but notes the large number of variables makes it difficult to predict how large numbers of people will react in a given situation. "A lot of crowd modeling is based on computational social science," Manocha says, noting both individual personalities (aggressive versus passive) and reactions (calm versus panic), along with cultural factors, play a large role in how people react to crowding.

"People in some parts of Asia are used to having small personal space around them," Manocha, a native of India, explains, "whereas in the Western cultures, pedestrians like to have more personal space. This results in much slower movement speeds as pedestrians slow down to have a larger personal space around them."

The result of this slowdown is an increase in crowd density, which can contribute to crushes, particularly if someone halts or falls. Manocha, whose group has been working with the Hajj Research Institute of Saudi Arabia to study and prevent crowd crushes, says that during the September 2015 Hajj disaster, which drew more than two million pilgrims to the Kaaba (a building at the center of Islam’s holiest site), crowd density reached between seven and nine persons per square meter, creating an extremely dangerous situation.

Manocha’s team has developed an algorithm to model density-dependent behaviors in crowd situations, which takes into account factors such as individuals’ stride length, walking speed, the size of the crowd, and factors such as the acceptable personal space required by the culture. The results are refined constantly by comparing this simulation to video of actual crowd behavior at real-world sites.

Still, it is very difficult to devise a standard crowd-behavior model since humans are, by their nature, somewhat unpredictable when unforeseen events come into play. "There’s very little data on how people actually react when a real panic situation arises," Manocha says, although he allows as more video of crowd behavior from disasters is captured on cellphone videos, it will be easier to validate and modify future models.

Another challenge with modeling crowd behavior for public safety lies in creating a useful visualization to query the data produced by the algorithm, according to Xinyue Ye, founding director of the Computational Social Science Lab at Kent State University.  Ye says creating a visualization to help query the data "becomes extremely important, because with a personal handheld computer, we cannot handle large amounts of data. But these [smart devices] are the ones that are useful for law enforcement agencies," which often need to make decisions on the fly during crowd actions like moving protests, celebrations, or mass evacuations.

One additional variable that has come into play in recent years is the use of user-generated search queries to identify, monitor, and predict how crowds will move. In response to the Shanghai Stampede, Baidu, the Chinese online mapping and search provider, has correlated user map search query data with mobile device positioning data to devise a model for measuring the risk of potential crowd disasters, and then invoked warnings about locations subject to potential overcrowding, one to three hours in advance.

"We found that the query data from Baidu Maps was strongly correlated with the location data, so from that, it’s a good predictor of how many people might be going to a specific area," explains Haishan Wu, a data scientist at Baidu, Inc. "We validated the correlation between the map queries and location data, and found it is strongly correlated.  On average, the correlation value is about 0.8."

Wu says future algorithms will have even more sources, such as social media postings, from which to pull real-time data on crowd behavior. "In the future, we can probably incorporate data from Weibo [the Chinese version of Twitter] into our algorithm for real-time learning," he says, noting these postings can also be correlated with map queries, standard Internet searches, and real-time location data to provide more information on crowd activity and movement. This tactic has already been tested in Milan, Italy in 2013, with accurate estimates of the number of people in a given location at a given time able to be extrapolated from mobile phone or Twitter data.

It is not only scientists using real-time data to predict crowd movement. In 2012, the New York City Police Department and Microsoft Corp. jointly developed the Domain Awareness System, which aggregates and analyzes in real-time public safety data provided by 911 calls, cameras, and mapping tools, to help law enforcement combat crime and protect citizens. According to Richard Zak, director of Justice and Public Safety Solutions at Microsoft, the system can also be configured to monitor and predict crowd movement. 

Beyond New York City, similar solutions have been deployed for law enforcement agencies in U.S. cities such as Oakland, CA, and Washington, D.C., as well as in Singapore, and cities in Brazil, Canada, and Qatar.

Keith Kirkpatrick is principal of 4K Research & Consulting, LLC, based in Lynbrook, NY.

Join the Discussion (0)

Become a Member or Sign In to Post a Comment

The Latest from CACM

Shape the Future of Computing

ACM encourages its members to take a direct hand in shaping the future of the association. There are more ways than ever to get involved.

Get Involved

Communications of the ACM (CACM) is now a fully Open Access publication.

By opening CACM to the world, we hope to increase engagement among the broader computer science community and encourage non-members to discover the rich resources ACM has to offer.

Learn More