Scientists build drones capable of detecting violence in crowds

But the algorithm gets less accurate the more people it tries to track

Scientists have trained drones to recognise violent behaviour in crowds using AI.

In a paper called Eye in the Sky, researchers from Cambridge University and India's technology and sciences institutes detailed how they fed an alogirthm videos of human poses to help their camera-fitted drones detect people committing violent acts.

The researchers claim the system boasts a 94% accuracy rate at identifying violent poses,and works in three steps: first the AI detects humans from aerial images, then it uses a system called "ScatterNet Hybrid Deep Learning" to interpret the pose of each detected human and finally the orientation of the limbs in the estimated pose are numbered and joined up like a coloured skeleton to identify individuals.

The algorithm used by the AI is trained to match five poses the researchers have deemed violent, which are classified as strangling, punching, kicking, shooting and stabbing.

Volunteers acted out the poses to train the AI, but they were generously spaced out and used exaggerated movements whilst acting out attacks. The report explains that the larger the crowd, and the more violent individuals within it, the less accurate the AI becomes.

"The accuracy of the Drone Surveillance System (DSS) decreases with the increase in the number of humans in the aerial image. This can be due to the inability of the FPN network to locate all the humans or the incapability of the SHDL network to estimate the pose of the humans accurately," the researchers wrote. "The incorrect pose can result in a wrong orientation vector which can lead the SVM to classify the activities incorrectly."

When one violent individual is in the crowd, the system is 94.1% accurate, which reduces to 90.6% with two, down to 88.3% for three, 87.8% for 4 and 84% for five violent individuals. By those figures, tracking violence in widespread incidents like the 2011 riots would currently be unworkable.

AI-powered recognition software is already in use among law enforcement bodies, despite fears that it is not accurate enough.

Both the Metropolitan Police and South Wales Police were accused of using dangerously inaccurate facial recognition technology by privacy campaign groups last month.

The groups revealed the Met had a failure rate of 98% when using facial recognition to identify suspects at last year's Notting Hill Carnival and that South Wales Police misidentified 2,400 innocent people and stored their information without their knowledge.

Image: Shutterstock

Featured Resources

Choosing a collaboration platform

Eight questions every IT leader should ask

Download now

Performance benchmark: PostgreSQL/ MongoDB

Helping developers choose a database

Download now

Customer service vs. customer experience

Three-step guide to modern customer experience

Download now

Taking a proactive approach to cyber security

A complete guide to penetration testing

Download now

Recommended

10 keys to AI success in 2021
Whitepaper

10 keys to AI success in 2021

10 Mar 2021
MLOps 101: The foundation for your AI strategy
Whitepaper

MLOps 101: The foundation for your AI strategy

10 Mar 2021
How to choose an AI vendor
Whitepaper

How to choose an AI vendor

10 Mar 2021
Realising the benefits of automated machine learning
Whitepaper

Realising the benefits of automated machine learning

10 Mar 2021

Most Popular

Microsoft is submerging servers in boiling liquid to prevent Teams outages
data centres

Microsoft is submerging servers in boiling liquid to prevent Teams outages

7 Apr 2021
How to find RAM speed, size and type
Laptops

How to find RAM speed, size and type

8 Apr 2021
REvil threatens to release Apple’s hardware schematics
ransomware

REvil threatens to release Apple’s hardware schematics

21 Apr 2021