![]()
Visiting guest. Why not sign in?
|
|
Predicting Crimes Cameras in Preventive Mode | |
Predicting Crimes
The Independent has an interesting article about Robot Cameras that 'Will Predict Crimes Before They Happen'. These smart CCTV systems that can monitor and learn patterns in movement. Simple novelty detection can be then performed, by comparing the current movement with ones that have been flagged OK! If the difference is too big, the computer alerts security guards, or the police, as something fishy is about to happen. The great thing about this is that very rudimentary computer vision algorithms are required to extract people from images, which means they can be reliably followed. The same technology has been successfully tested in the London underground, to check for overcrowding and suicide attempts! All these problems are both interesting and useful to solve -- far more than face recognition based approaches... |
|
Novelty detection
When rodney was only a two axis robot - similar to a security camera - I tried something similar to this. For any particular direction of the camera you can work out an average image. The amount of novelty can then be worked out by comparing the current image to the expected average. It looks like they've used a similar approach either using average visual motion or using markov models to predict patterns of motion change. Anything which then deviates significantly from the average patterns can then be marked as interesting. Going one step further you can use this sort of novelty detection to drive a learning system. Generally in reinforcement learning systems its important to learn only about events which are significantly novel. If a sizable reinforcement signal is generated continuously then the learning tends to be poor, with the system being continually dragged in one direction or another. - Bob |
|
Images & Vectors
I sincerely believe they use a filter that can extract blobs over time. If you factor out the expected image, and increase the contrast you can then end up with just a point in projected 2D space. You could theoretically then project that point back into 3D, but I don't believe that's necessary. Using that 2D point, things are very simple. I think your average case would involve walking straight through the area at a constant pace. If the pedestrian slows down and starts to loiter with a purpose, or if a group of blobs forms around the main entrance, then something fishy is going on -- so you signal the security guard. You could learn this online, as you're just gathering trends. This is probably all done with rules, as I can't see a way of getting markov models to do all of this (especially the social gathering bit). Interesting discussion though ;) |
|
Surveillance Cameras to Predict Behaviour
BBC News has a good article on the same topic, as Surveillance Cameras to Predict Behaviour. They mention the cameras in the London underground used for predicting suicides... it's interesting how they can tell just by the way someone moves! Even a human watching often couldn't predict the accident until too late. |
|