Surveillance camera checks out 36 million faces in a single second
When it comes to attaching a name to a face, I am really, really bad at that. However, when you get a computer to do so, you can be sure that the sophisticated algorithms used these days (especially those in the movies) are more than capable of detecting just the face you were looking for in a jiffy. I do have a question though – how come the software that government agencies and intelligence outfits use in the movies tend to be out of this world? From afar, they are able to capture a face, blow it up a gazillion times and even add more information to the pixels already there for a crystal clear mugshot.
Try asking the authorities to do that in real life when it comes to CCTV footage of a thief from a hundred yards or so, and they will say that it is nigh impossible, as all you will get is a blurred picture of a face that could very well resembles a million other people out there on the streets. Well, here is a snippet of good news for you – there might be hope just yet, although this is just a small step. Hitachi Kokusai Electric has come up with a surveillance camera system that is capable of searching through data on 36 million faces – and the kicker is this, it is capable of doing so in a single second.
This particular system will be able to automatically detect a face from either surveillance footage or a regular photo, and search for it. All search results will be instantaneously displayed, showing off thumbnail images of potential candidates. Whenever a thumbnail is chosen, the associated recorded surveillance footage can be viewed for one to go through the person’s actions prior to and after the image was taken.
According to Hitachi, “This high speed is achieved by detecting faces through image recognition when the footage from the camera is recorded, and also by grouping similar faces.” It will assume that faces are turning within around 30 degrees in the horizontal and vertical directions from the camera, and the faces are at least 40 x 40 pixels in size.