Public events are a numbers game. If a hundred people show up at a protest in Washington, D.C., who cares? But a hundred thousand? Okay, let’s hear what they have to say. With a country fiercely split politically, each side will be bringing out as many supporters as they can today and tomorrow: conservatives with the inauguration of President-elect Donald Trump on Friday, and liberals with the Women’s March on Washington on Saturday. The high estimate for the Inauguration is 800,000 people, while 215,000 people have registered for the Women’s March on the event’s Facebook page. (There are also Women’s Marches planned for other major U.S. cities.) Each side will have an incentive to exaggerate their turnout, and expect plenty of heated debate about the numbers this weekend. But technology is making it harder for them to stretch the truth.
We can expect especially accurate estimates for the Women’s March in D.C., though the best count probably won’t be out until several days later. Unlike Friday’s Inauguration, which is a no-fly zone, the Saturday protest will be photographed from the air. At least one firm, Digital Design and Imaging Service (DDIS), will do a tally based on a combination of high-tech tools and grunt work.
The company gained notoriety (and the ire of organizers) for counting Glenn Beck’s August 2010 Restoring Honor rally at the Lincoln Memorial. NBC estimated 300,000 attendees, and Beck said there might have been up to 650,000. CBS, which hired DDIS to do a count, came up with 80,000 using a pretty thorough method.
For that rally, DDIS surveyed the National Mall before the event and produced a 3D map dividing it into different sections; it also estimated where people are most likely to congregate (like by the stage or under trees in hot, sunny weather). The company flew a tethered aerostat balloon carrying a nine-camera array a few hundred feet over the event to capture 360-degree panoramic images at various heights (to get different angles on the crowd).
Then came the grunt work: Counting heads in the photos. Fortunately, they didn’t have to count every single person. Rather, DDIS categorized the grids by how densely people were packed. If there were 500 people in one grid, that count was multiplied by the number of grids that it estimated to be equally dense with people. DDIS used the same method for the Steven Colbert-John Stewart Rally to Restore Sanity and/or Fear in October 2010 (estimated at 87,000).
It will fly again over the Women’s March on January 21, providing counts for two media organizations, says company president Curt Westergard, without naming the clients. The technology hasn’t changed much since 2010.
“Our 3D computer-generated, hand-counted polygon system is pretty well perfected. No evolution needed on that,” boasts Westergard in an email. “Any advance would be airspace politics,” he adds, noting that he won’t be able to fly over the inauguration with a balloon (or drone). Westergard did fly over the site of Barack Obama’s first inauguration, in 2009, but for security reasons (outgoing President Bush arrived earlier than expected) had to bring the balloon down long before the event started and the full crowd materialized.
Satellites will by flying over Trump’s inauguration, as they did in 2009. Journalism professor Stephen Doig, one of the gurus of crowd counting, used an image from satellite company GeoEye to estimate a crowd of 800,000 people for Obama’s first Inauguration. Official estimates for the event, though, were over twice as high, at 1.8 million, which may actually be impossible, say researchers. DDIS modeled how big a very tightly packed crowd of a million people would stretch back from the Capitol building along the Mall: It would extend across the Potomac River into Virginia.
Attendance at Trump’s inauguration is forecast to be up to 800,000, according to the U.S. Armed Forces Joint Task Force, which supports the ceremony. When final numbers come in, the 800,000 estimated for Trump may also prove wildly optimistic.
DigitalGlobe, which bought GeoEye in 2013, will be releasing photos of the inauguration, as well as the Women’s March; other satellite firms will likely be taking pictures, too. There are some drawbacks for shots from so high up—nearly 400 miles versus about 700 feet for a balloon.
“Although satellite imagery is a very good resource for crowd counting, using only satellite imagery is not going to be the most accurate method,” says Charlie Loyd, imagery engineer at Mapbox, a mapping company that uses satellite imagery from DigitalGlobe. Its new satellites can capture details down to a resolution of about a foot (30 centimeters) across. “That’s on the edge of what works reasonably well for crowd size estimation,” says Loyd, in an email. “You still get problems like people standing in each other’s shadows and therefore not being fully resolved.” However, it’s an improvement from the 20-inches (50cm) resolution in the 2009 photos.
Cutting-edge satellites, 3D modeling, and hand-counting people’s heads make for an incongruous combination. The artificial intelligence wave that’s sweeping across what seems like every business is finally coming to crowd estimates. In 2015, researchers at the University of Central Florida (UCF) debuted algorithms that proved to be as accurate as hand counting, and a hell of a lot faster. The first big test came in counting the turnout at a September, 2015 pro-independence rally for the Catalonia region of Spain. Researchers at the Pompeu Fabra University in Barcelona provided 67 photos that together covered the entire crowd. Within a half hour, the Central Florida team had a tally: 530,000. That might be big by U.S. protest standards, but it’s considerably less than the estimates given by the Catalonia organizers.
The Barcelona academics checked the results by hand-counting some of the images. “They were fairly close to each other, so that’s why they didn’t bother manually counting for all of them,” says Haroon Idrees of the UCF team. Their system works on several levels, all based on machine learning—analyzing many photos of crowds to recognize how people may appear. If the crowd is small enough, like a few thousand, the system can often pick out individual’s heads from an areal view. “You need 15 x 15 [pixels per person] or slightly more than that, to get good accuracy,” says Idrees. This is where the accuracy roughly matches what a human can do, but it’s far faster.
UCF’s system can keep counting when the images are too poor for people to make sense of. “Even if the humans are not able to count by hand, our algorithms still give good results, because we can train our algorithms based on some blurred images,” says Idrees. “The accuracy may go down slightly but still the accuracy will be better than human counting.” The UCF effort can go even farther, sampling patterns of light and dark pixels, for instance, to make a guess at how many people are present in a photo where nothing that a human eye would recognize as a person is present.
UCF’s system probably won’t be used for the Inauguration or Women’s march. The university provides the technology for crowd counting, but only works on specific projects when other people, like the Barcelona researchers, approach them. “Regarding the anticipated Washington crowds, we haven’t been contacted by anyone so far,” says Idrees.
The university’s biggest collaboration right now is with a Saudi Arabian organization that counts crowds of worshippers visiting Mecca. The annual Haj pilgrimage suffers from unwieldy crowds that often break into fatal stampedes. It’s one major public event where organizers are actually hoping and praying for a smaller turnout.