IEEE Future Directions
Related Articles
Fifteen years ago at the University of Southern California a team of researchers developed a program, Mugspot, that was able to recognise faces in images taken by a security camera and compare that face to a data base of faces with the aim of identifying a specific person. The software could work around moustaches, different hair style – color, to make a match.
That program came back to my mind as I read a news from the University of Washington reporting on a new program that let security cameras talk to one another and follow people as they move out of their field of vision.
Notice that it is not just a matter of recognising the same pattern in images taken by different cameras. Each camera has its own perspective and a person that appear in from of one my be seen from the back in another or from up to down and everything in between. The algorithm has to work out what our brain does constantly. We recognise our friend independently (within certain ranges of course) of his position with respect to us.
Image recognition and more specifically recognising the identity of a person is really tricky. This is why after fifteen years from that first experiment with Mugspot we are still struggling with identification.
To achieve a good recognition the researchers have the cameras "talking" with one another and first having a training period of a few minutes to "understand" their different point of views with respect to a fixed panorama. Once they compute the differences they start applying those differences in the capturing and matching of moving people.
They have made a number of experiments in the university campus, connecting cameras from fixed positions, camera mounted on cars and even cameras on drones demonstrating the capability to identify a person and follow his movements as he moves from camera to camera.ì, recapturing him after he has disappeared in "shadows" areas.
The researchers imagine several applications scenario, including, and this was sort of a surprise to me, tracking people behaviour in large department stores and converting this information into marketing opportunities, like understanding what that person may like and be undecided upon and sending him a discount coupon to trigger an impulse buying.
Well this is, I guess, in the line of smart street retail one of the EIT ICT Labs High Impact Initiatives, although I hope that privacy concerns are better taken into account!
