4

Inspired by this discussion about recognizing human actions, I have found the Fall-Detection project which detects humans falling on the ground from a CCTV camera feed, and which can consider alerting the hospital authorities.

My question is, are there any existing real-life implementations or research projects which specifically use live video feed from the surveillance cameras in order to detect crime using convnets (or similar approaches)? If so, how do they work, briefly? Do they automatically inform the police about the crime with the details what happened and where?

For example car accidents, physical assaults, robberies, violent disturbances, weapon attacks, etc.

Community
  • 1
kenorb
  • 10,423
  • 3
  • 43
  • 91
  • 2
    Automatic crime/accident detection (using conv nets or otherwise) has been around for decades [e.g. here](http://ieeexplore.ieee.org/xpl/login.jsp?tp=&arnumber=491539&url=http%3A%2F%2Fieeexplore.ieee.org%2Fxpls%2Fabs_all.jsp%3Farnumber%3D491539) The problem isn't so much whether a CNN could be built to detect crime, but the moral and liability issues surrounding leaving the decision whether or not to send emergency services to a crime to a computer. That's what restricts all existing systems to only alert an operator to an incident, and let the humans decide what to do. –  Aug 21 '16 at 19:17

1 Answers1

4

After a bit of research I found something kind of close:

They, however, makes no mention of what specific methods they use.

So a crime detection system as that is written does not exist, but abnormal behaviour detection systems do.

An accurate generalized system seems intuitively infeasible, however. Commiting a crime, unlike falling, is a complex behavior, and takes so many forms. A camera watching a store's counter like at a 7-11 could perhaps see that the 'customer''s arm is strangely reaching across the counter, and the attendant is moving a lot more than usual suddenly, but aside from very specific cases like this such a system is currently quite unfeasible.

Crimes are unusual, relatively speaking, and their dramatic nature means that even the simplest crimes play out in very different ways. Perhaps you could in this case you could try to look for images of a gun, or someone with their hands up. So, looking for unusual, detectable behavioural mannerisms may be possible, but not crime detection.

Ultimately, while you may be able to make (possibly pretty good) systems to detect specific crimes in specific environments, that's all we got for now.

P.S. - Do these camera's also get audio signals? That is also an interesting facet to consider ("PUT YOUR HANDS UP / GIVE ME ALL YOUR MONEY")

kenorb
  • 10,423
  • 3
  • 43
  • 91
Avik Mohan
  • 706
  • 3
  • 12
  • The "audio signals" approach was in fact done earlier than the image processing. There are public systems that uses audio to detect screams of distress, as well as trigger phrases, and automatically point and zoom the nearest CCTV camera to the source. See [here](http://news.bbc.co.uk/1/hi/scotland/7886656.stm). Relatively speaking, it's a lot easier to detect changes in voice tone associated with violence than it is image processing. The decision making is still left to humans though, the system alerts a human operator to a suspicious event, with the operator then deciding what action to take –  Aug 21 '16 at 19:05