The system works by analyzing live video feeds with a powerful combination of deep learning models, including Scaled-YOLOv4 and EfficientDet. It’s fast (capable of spotting fire in as little as 0.016 seconds per frame) and accurate, thanks to a smart design that uses multiple models working together to reduce false alarms. Best of all, it doesn’t require expensive thermal cameras or special sensors. If a building has security cameras and internet access, it could use this system.
Why This Matters
Fires move fast. Today’s homes and buildings, built with synthetic materials and open floor plans, can go from spark to flashover in minutes. That leaves very little time to react. And yet, most fire detection systems still rely on traditional smoke detectors, which only trigger an alarm once a certain level of smoke hits them directly.
This means delays in detection, and often, tragedy. Each year, fires cause nearly 3700 deaths in the US, with over $23 billion in property damage. In 11 % of fatal residential fires, detectors were either missing or didn’t work. Clearly, there's room for improvement.
Surveillance cameras, on the other hand, are nearly everywhere. What’s been missing is a smart, reliable way to use those video feeds to detect fires early without constantly triggering false alarms over a sunset, a bonfire on TV, or a car with bright red paint.
How the System Works
The NYU team designed the system as a lightweight, cloud-based IoT architecture that fits into three layers:
- Perception Layer: This is where the video comes in. Regular CCTV or IP cameras act as passive eyes, sending unprocessed video—no fancy hardware or onboard AI needed.
- Network Layer: The footage is streamed over the internet using standard IoT protocols.
- Application Layer: All the real work happens here. Cloud servers (like AWS EC2) run powerful AI models that analyze the video in real-time, looking for signs of fire or smoke.
The team tested several top-tier object detection models like Faster-RCNN, EfficientDet, and multiple YOLO variants to see which worked best. Scaled-YOLOv4 came out on top with an average precision of 80.6% and lightning-fast processing at 0.016 seconds per frame. EfficientDet followed closely. Slower models like Faster-RCNN didn’t make the cut due to longer processing times and lower accuracy.
Keeping False Alarms in Check
False alarms are a major concern with any automated detection system, especially one that deals with emergencies. To make sure the AI doesn’t cry wolf every time it sees something vaguely flame-colored, the researchers added two key features:
- Multi-Model Agreement: Instead of relying on just one AI model, the system requires multiple models to agree before raising an alert. This makes it much less likely to get fooled by a red car, sunset, or even a fake flame image.
- Temporal Tracking: It’s not just about what’s in the frame; it’s about how it moves. The system tracks detected smoke or fire across multiple frames, watching how it changes over time. A real fire moves and grows; a static image doesn’t. That difference helps the AI tell real threats from false ones.
When a fire is confirmed, the system automatically stores video clips in AWS S3 and sends alerts via email and text using AWS’s notification services (SNS and SES). Emergency responders get notified instantly, and with confidence that it’s not just a false alarm.
Final Thoughts
This research doesn’t just offer a smarter way to detect fires; it rethinks how we use the technology already around us. Security cameras, once passive observers, can now actively support life-saving decisions. With minimal setup and no need for expensive upgrades, the system lowers the barrier to real-time fire detection in places that have historically been difficult to monitor.
As the climate crisis brings more extreme fire risks and urban environments grow more complex, tools like this will be critical. Early detection saves lives, limits damage, and gives responders a fighting chance. And now, with AI and cloud infrastructure working together, that kind of protection is more accessible than ever.
Journal Reference
Panindre, P., Acharya, S., Kalidindi, N., & Kumar, S. (2025). Artificial Intelligence-Integrated Autonomous IoT Alert System for Real-Time Remote Fire and Smoke Detection in Live Video Streams. IEEE Internet of Things Journal, 1–1. DOI:10.1109/jiot.2025.3598979. https://ieeexplore.ieee.org/abstract/document/11127189
Disclaimer: The views expressed here are those of the author expressed in their private capacity and do not necessarily represent the views of AZoM.com Limited T/A AZoNetwork the owner and operator of this website. This disclaimer forms part of the Terms and conditions of use of this website.