Even though it dates me as someone who has been around a while, the first time I can remember the “face in the crowd” phenomenon coming to notice was in the case of Arthur Bremer. Bremer was the man who tried to assassinate then-presidential candidate George Wallace in 1972, while Wallace was speaking at a rally in Laurel (Md.). He didn’t succeed in killing Wallace, but did make him a paraplegic, confining him to a wheelchair for the rest of his life.
In the investigation that followed the assassination attempt, the U.S. Secret Service found Bremer in photographs of crowds in attendance at other Wallace rallies around the country. He was usually standing with his hands in his jacket pockets, wearing sunglasses, and smiling in a creepy way. Even before that, arson investigators knew the value of photographing the crowds of spectators at building fires. Arsonists often return to the scene of the crime to watch the results of their handiwork. If the same person shows up at multiple fire scenes, it’s probably a good idea to try and locate them for a chat.
Researchers Kevin Bowyer and Patrick Flynn in the Department of Computer Science and Engineering at Notre Dame University got interested in this problem when Bowyer attended a conference that included a discussion on counterterrorism. One of the conference participants mentioned the problem of scanning videos or photographs of crowd scenes in the aftermath of a detonation of an improvised explosive device (IED) or suicide bomb. There could be hundreds of people in the pictures, and it was close to impossible to compare all of the faces with all of the others, looking for a commonality.
The new technology uses the geometry of certain unalterable facial features to identify each face in the image frame and compare it to faces from other images. While someone can wear a disguise that changes their hairstyle, facial hair or even skin tone, the distance between the pupils, and the angle that connecting line forms with the tip of the nose, among other features, is always consistent. The geometry of each face isn’t unique, but is sufficiently distinctive that is greatly reduces the number of faces a human has to compare to something manageable. The software is called the Questionable Observer Detector, or QuOD.
Just identifying the faces in an image used to be quite a feat for a computer, but that technology is now so “ho-hum” that it’s included in consumer-grade point-and-shoot cameras. For a couple of hundred bucks, you can buy a camera that will find every face in a group and snap the picture only when everyone has their eyes open and is smiling.
While this is a great development that you can see through a short video on the Notre Dame website and below, there are still some problems to work out before it can be put to work. Video clips are of highly varying quality and resolution, and some don’t have enough detail to identify the faces. Lighting and clothing can put a face in shadow, and not everyone may be looking directly into the camera. All that adds complexity to the identification task, which also increases with the number of videos to be analyzed.
One refreshing aspect to this technology is that civil rights and privacy advocates are less likely to find fault with it than with facial recognition systems that compare every face detected to a database. Critics say this casts too wide a net, increasing the chances for false positives and detentions of people who haven’t done anything but look similar to someone who is wanted. When the comparison database is limited to people who showed up to watch bodies being recovered from a bombing, there is less of an opportunity for error.
