NYU Tandon School of Engineering: Alarm triggered in less than a second
New image analysis software from NYU Tandon School of Engineering (https://engineering.nyu.edu/ ) uses footage from surveillance cameras on streets and in buildings to detect fires at an early stage. The developers believe that this could save thousands of lives and billions of pounds. According to researchers Prabodh Panindre and Sunil Kumar, almost 3,700 people die in fires every year in the United States alone. The damage amounts to £23 billion.
A little smoke is enough
‘A camera can monitor a much larger area than conventional fire detectors, and we can detect fires in their early stages when they produce so little smoke that conventional systems do not respond,’ says Pandindre. The video footage is analysed using artificial intelligence (AI) software that detects fires within 0.016 seconds, gaining additional minutes for evacuation and emergency measures.
The software combines several AI algorithms. Instead of relying on a single AI model that could confuse a red car or a sunset with a fire, the system requires multiple indicators to match before reporting a fire, significantly reducing false alarms that would undermine the credibility of such an alarm system.
By analysing changes in size and shape in successive video images of areas with initial suspicion of fire, the algorithm distinguishes between a real, spreading fire and a static image of flames on a wall. “Real fires are dynamic, spread and change shape. Our system tracks these changes over time and achieves 92.6 per cent accuracy in eliminating false alarms,” said Kumar.
Central server analyses
The researchers trained the software by creating a comprehensive custom image dataset covering all five fire classes recognised by the National Fire Protection Association (https://www.nfpa.org/ ) – from common combustible materials to cable fires and cooking accidents. According to the company, the system achieved an 80.6 per cent accuracy rate in tests.
Standard surveillance cameras send raw video footage to a central server, which analyses it. If the server detects a fire, video clips are automatically generated and real-time alerts are sent via email and text message. It also works with mobile surveillance cameras in drones or other unmanned aerial vehicles that can be used to detect forest fires.
