COULD ALGORITHMS HELP DETECT VIOLENCE?
Drones may one day hover over crowds in public spaces and at large events, scanning for signs of violent behaviour, according to a team of researchers from the UK and India.
The researchers developed a computer algorithm that maps human postures and attempts to match them to ‘violent’ archetypes: strangling, punching, kicking, shooting and stabbing. The team then trained this algorithm on recorded footage of fake fights and once it had become accurate they tested out on a live video of a group of actors to see whether the tech would work in real-time. The accuracy of the system wasn’t perfect. The researchers reported 94 per cent accuracy when it came to identifying ‘violent’ poses, which dropped depending on how many people were in the frame. More importantly, this new technology was tested on between 2 and 10 volunteers generously spaced apart who were clearly pretending to attack one another – not exactly a real-world situation. The researchers plan to further test the software at two upcoming festivals in India.
Both Google and Facebook have published studies showing that neural networks can be used to track poses, but it’s hard to imagine the technology being deployed soon. The images the system analysed were taken from between two and eight metres away, and given how loud drones are, festival goers probably won’t appreciate the noise, even if it is keeping them safe.