Computer vision algorithms pick out petty crime in CCTV footage

[neonix]

Computer vision algorithms pick out petty crime in CCTV footage
New Scientist 4 January 2017
By Timothy Revell

https://www.newscientist.com/article/2116970-computer-vision-algorithms-pick-out-petty-crime-in-cctv-footage/

Petty criminals had better watch out. A computer vision system has been developed that detects suspicious behaviour in CCTV footage as it happens. The system can then alert CCTV operators to intervene, and save the footage in case it is needed for evidence.

Researchers involved in the P-REACT project project, which is the work of a consortium of European companies and organisations and is partly funded by a grant from the European Commission, say the surveillance technology could help catch criminals in the act and relieve police of “digital evidence overload” by highlighting video clips most likely to be relevant to investigations.

“If a camera at a gas station picks up suspicious activity, the video footage will be sent to the cloud, people at the gas station will be alerted, and nearby cameras will be told to look out for the criminals too,” says project coordinator Juan Irujo at Vicomtech, a research foundation in San Sebastian, Spain.

P-REACT tracks people’s movements to work out whether they’re simply walking along a street, for instance, or doing something dodgy. Its algorithms have been trained on sample scenes of people fighting, chasing someone or snatching a bag. They had to be finely tuned to identify these activities: hugging can look a lot like fighting, for example, while running can be mistaken for giving chase.

Petty criminals

The first tests of P-REACT were run last year in a carefully controlled environment, with actors playing the roles of petty criminals. In this idealised set-up the system reacted flawlessly, catching every play-acted crime – though it is likely to have more difficulty in a real-world scenario where scenes are less predictable. The technology was presented at the International Conference on Imaging for Crime Detection and Prevention in November, with more extensive field trials planned for the future.

Meanwhile, aspects of the technology are already being rolled into tools for use by police.

Dublin-based company Kinesense is developing products that will give police access to CCTV clips selected by the P-REACT system. “One of the biggest challenges police face is digital evidence overload,” says chief technology officer Mark Sugrue. “P-REACT solves this by letting the camera send only important clips.”

Sean Gaines at Vicomtech says P-REACT could also help prevent profiling based on race or age, as the system only analyses movement and is not subject to the conscious or unconscious biases that might influence a CCTV operator’s decisions. “Our algorithms do not take appearance into account, only actions matter,” he says.

Teo de Campos at the University of Brazil says that systems like P-REACT are a useful tool that “help the operator in a surveillance room to focus on relevant cameras or even in relevant regions of images”. Similar systems are probably already in use in some places, he says, but the details are often kept secret.

However, although systems like P-REACT can help pick out unusual behaviour, Marcos Nieto at Vicomtech emphasises that they can’t actually tell if an act is criminal. That particular job should be “left for the human beings.”

 

[Gawan]

Thanks for sharing. I think Orwell turns again in his grave if he could read this. And supposedly it is all for our safety. The cameras would need as well to have good quality imo, otherwise the software misses important pixels or misreads them.

 

[williamsj]

This type of monitoring is not new, research was underway in detecting suspicious behaviour / movements in the the 1990's.

See: http://journals.sagepub.com/doi/abs/10.1068/p3402

The whole point here is that it does not require detailed video, it rather detects a person or car for that matter as a block and tracks their movements looking for specific patterns of behaviour.

For those interested, there are numerous examples on You-Tube from several Universities etc.

At the other end of the spectrum we have:

Abstract: http://ieeexplore.ieee.org/document/6836019/
One of the biggest difficulties in human action analysis is the temporal complexity and structure of actions. By breaking actions down into smaller temporal pieces, it may be possible to enhance action recognition, or allow unsupervised temporal action clustering. We propose a temporal segmentation system for human action recognition based on person tracking and a novel segmentation algorithm.

Big brother has been watching us for years, they just don't tell us...