Smart CCTV learns to spot suspicious types

the difference between a suicide bomber and a cleaner? It sounds like
the opening line of a sick joke, but for computer scientists working on
intelligent video-surveillance software, being able to make that
distinction is a key goal.

[] Current CCTV systems can collect masses of data, but little of it is used, says Shaogang Gong,
a computer-vision computation researcher at Queen Mary, University of
London. "What we really need are better ways to mine that data," he

Gong is leading an international team of researchers to develop a next-generation CCTV system, called Samurai,
which is capable of identifying and tracking individuals that act
suspiciously in crowded public spaces. It uses algorithms to profile
people’s behaviour, learning about how people usually behave in the
environments where it is deployed. It can also take changes in lighting
conditions into account, enabling it to track people as they move from
one camera’s viewing field to another.

improve the tracking of an individual at an airport, the system can
also learn the routes people are likely to take – straight from the
entrance to check-in, say. It can even follow a target as they move in
a crowd, using the characteristic shape of the person, their luggage
and the people they are walking with, to follow them as they walk
between different camera views.

is designed to issue alerts when it detects behaviour that differs from
the norm, and adjusts its reasoning based on feedback. So an operator
might reassure the system that the person with a mop appearing to
loiter in a busy thoroughfare is no threat. When another person with a
mop exhibits similar behaviour, it will remember that this is not a
situation that needs flagging up.

video analysis tools already exist, they tend to operate according to
rigid, predefined rules, says Gong, and cannot follow a large number of
people across multiple cameras situated in busy public spaces.

Samurai team last month demonstrated the system to commercial partners
including BAA Airports in the UK. The researchers claim the prototype
system successfully identified potential threats which may have been
missed by human operators, using footage collected at Heathrow airport.
The Samurai team has funding to continue refining their software until
the end of 2011.

"The use of relevant feedback from human operators will be a very important part of these technologies," says Paul Miller,
of Queen’s University’s Centre for Secure Information Technologies in
Belfast, UK, who is leading a project to develop a video-analysis
system capable of predicting assaults on buses. "The key is developing
learning algorithms that work not only in the lab but that are robust
in real-world applications."