TURNING a script into a film is a Hollywood speciality but the “inverse
Hollywood problem” of turning a film into a script is a lot harder, especially
for a computer. But now there’s a system that can analyse video footage and
generate a script that describes what the people in it are doing. It should
appeal to employers who want to track the activities of their staff, legal or
otherwise, in offices and other workplaces.
Computer systems can already recognise images of objects and people, but they
find it much tougher to classify what people are doing. From a video of someone
picking up a briefcase, say, a computer may be able to recognise the briefcase
and the person. But the fact that the human is holding a briefcase is much more
difficult to detect.
Now Muburak Shah and Douglas Ayers at the University of Central Florida in
Orlando have developed a system that can tell whether an object has been grasped
and lifted, by tracking the trajectories of people’s hands and fingers.
Advertisement
With a single camera observing an office, the system can spot actions such as
making a phone call or opening cabinets and drawers. It also monitors whether
people carry objects out of the room. The computer then produces a script
describing these actions. “Analysing the trajectories of hands and fingers is
the hardest part, but our accuracy is close to 100 per cent,” says Shah.
A computer that has been given information about who owns which objects
should be able to spot when someone is using other people’s phones or computers,
or if they are stealing objects from the room.
The system is more invasive than current surveillance methods, says Barry
Steinhardt, an expert on computer privacy at the American Civil Liberties Union
in New York. Employers can already monitor e-mails, phone conversations and
computer keystrokes, but until now analysing video footage has been tedious and
time-consuming.
-
More at:
Image and Vision Computing (vol 19, p 833)