People love to take selfies, but it's a love that can prove fatal. A growing number of people die each year while snapping photos of themselves on cliffs, on railroad tracks and other hazardous spots. Researchers in Pittsburgh and in India are looking for ways to reduce this risk.

In a new study, researchers at Carnegie Mellon University and the Indraprastha Institute of Information Technology in Delhi (IIITD) scoured public records to compile a list of 127 deaths associated with selfies worldwide between March 2014 and September 2016. After analyzing those selfie deaths, they designed a system that uses location, image and text to classify whether a selfie was dangerous or not.

As soon as the study was posted to arXiv, a popular online repository of electronic preprints of scientific papers, news media took notice. In the flood of stories that were posted and broadcast, most focused on the hows and whys of such deaths, from young lovers in India struck by trains to gun lovers in the U.S. accidentally shooting themselves.

But one of the primary goals of the research, said Hemank Lamba, a Ph.D. student in CMU's Institute for Software Research, is to develop an app to prevent these deaths. Such an app, not yet developed, might warn users when they are preparing to take a hazardous selfie or might even temporarily disable a smartphone's selfie function.

Another goal would be to use this data to inform public policy, perhaps establishing "no selfie zones" in hazardous areas or creating public education programs to help people recognize dangerous selfie practices, said Megha Arora, a master's student in CMU's Computer Science Department.

Ponnurangam Kumaraguru, an associate professor at IIITD who earned his Ph.D. in computer science at CMU in 2009, said he launched the study last summer after a report of a selfie death circulated on his research group's mailing list.

"It disturbed me quite a bit; every life is precious," Kumaraguru said. "As we looked into it, we realized not a lot of technical work had been done in this area."

To find a solution, the researchers — including Lamba and Arora, both former students of Kumaraguru at IIITD — began collecting news reports on selfie-related deaths from around the world to help them characterize the types of selfies that often lead to deaths.

The most common cause was from people falling from buildings, cliffs and other high places. Water-related selfies accounted for the most group deaths, often from capsized boats. In India, a number of deaths occurred when friends or lovers posed on railroad tracks, which is widely regarded as a symbol of long-term commitment in that culture. Gun-related deaths in selfies occurred only in the U.S. and Russia. Road- and vehicle-related selfies and animal-related selfies also were associated with deaths. (See the map below.)

Men accounted for three out of every four deaths, perhaps because men are more likely than women to seek out dangerous locations for selfies. Taking dangerous selfies has become a trend among some people online, who often tweet using hashtags such as #dangerousselfie and #extremeselfie.

Arora said the researchers were able to take advantage of those hashtags in the next phase of the study, assembling more than a thousand selfies taken in potentially dangerous situations. They developed a method to train a computer to recognize these situations. One factor is based on location — at extreme heights, near bodies of water, and near rail and roadways. Analysis of the image itself and of any text in the image also helped train the computer to classify a selfie as dangerous.

In testing this classifier against an annotated set of selfies, including 396 dangerous and 2,676 non-dangerous selfies, the system was able to tell the difference with 73 percent accuracy. That percentage takes into account both false positives and false negatives, Lamba said.

Creating an app to reduce selfie risk remains problematic, however. The researchers would like to improve the accuracy rate of their method, of course, but there also are questions about how such an app might be designed. Lamba said an app that simply identified dangerous situations might backfire — providing a bragging point for those users who want to take risky selfies.

"There can be no app for stupidity," is the way a lot of online commenters put it after the selfie death story went viral, Lamba said.

A properly designed app, however, would have applications beyond just selfies, Kumaraguru said. Augmented reality (AR) games, such as Pokemon Go, can sometimes put users at risk if they're not paying attention to the real world, he noted. With more AR games no doubt on the way, the potential risks for players could rise substantially.

Though this study of selfie deaths unites researchers around the globe, it nevertheless reflects a distinctive Carnegie Mellon approach, Kumaraguru said.

"When you see a problem in society," he explained, "you find ways to use technology to solve it."