Fire and smoke detection results raise questions for smart cities

Detection system combines two AI models

Researchers at Jouf University in Saudi Arabia have developed an artificial intelligence framework for early smoke and fire detection in smart city environments.

A study by Amr Abozeid and Rayan Alanazi describes a hybrid system that combines a Vision Transformer with the YOLOv8 object detection architecture.

The framework is designed to identify early-stage smoke and fire patterns in complex visual scenes and support faster response through smart city monitoring systems.

The study says conventional heat and smoke detectors often raise alerts only after smoke reaches the device.

It adds that existing video-based systems can also produce false alarms when visual conditions include dust, clouds or changes in lighting.

The Vision Transformer analyses global visual patterns in an image so the system can detect subtle smoke characteristics and spatial relationships across a scene.

The YOLOv8 module then performs real-time object detection to localise smoke and fire regions at speed.

Detection results and next steps

According to the study in Scientific Reports, the researchers trained and tested the model on more than 7,000 images from urban and rural datasets.

These images included smoke and fire scenes captured under different lighting conditions and in different environments.

The system achieved 99.2% accuracy, with precision of 98.5% and recall of 97.8%.

Its F1 score reached 98.1% and inference latency was about 45 milliseconds, which the study says enabled real-time detection at about 22 frames per second.

The researchers reported an accuracy increase of about 4.3% compared with conventional convolutional neural network approaches and other single-model detection systems.

The study also says the framework performed better than several existing detection models across precision, recall and localisation accuracy.

It states that the hybrid architecture can distinguish real smoke or fire from visually similar features such as printed flames or static images by analysing contextual and spatial characteristics rather than relying only on colour.

The authors say further validation in real-world settings is still needed, with future work set to examine thermal imaging and environmental sensor inputs for low-visibility conditions.