Blazing Trails in Fire Detection

Fires exact a devastating toll on human life and property, leaving behind a trail of destruction and heartache. The casualties and property damage caused by fires are staggering, with statistics painting a grim picture of the scope of this problem. According to the National Fire Protection Association, in the United States alone, there were an estimated 1.5 million fires reported in 2022, resulting in 3,790 civilian fire deaths, 13,250 civilian fire injuries, and $18 billion in direct property damage.

Despite advances in fire safety measures and firefighting techniques, many of these tragedies could be avoided or mitigated with better early detection techniques. Early detection is critical in preventing fires from escalating into uncontrollable infernos, allowing for timely intervention and evacuation efforts. Technologies such as smoke detectors, heat sensors, and advanced fire alarm systems play a crucial role in early detection and alerting authorities and residents to potential fire hazards.

Unfortunately, these existing technologies are far from perfect due to factors like limited detection ranges and low detection accuracy levels. As such, blazes can get out of control before building occupants are made aware of the situation, which might mean that the warning comes too late.

Technological advances in computer vision have led to new solutions that are not plagued by these problems of the past, offering new hope for early detection of fires. However, certain limitations have also hindered these new systems to date. Many of them struggle under low light conditions, and the detectors can be quite expensive because of the substantial computational resources they require.

A small team of engineers at the Zhengzhou University of Light Industry have just published a paper that may help to resolve many present issues with computer vision-based fire detection, however. They started with the YOLOv5 object detection algorithm, then enhanced it in a number of ways, to build a more accurate and efficient fire detection algorithm. This improved algorithm, called YOLOFM, increases average precision rates by as much as eight percent, which could be the difference between life and death in an emergency.

A dataset with over 18,000 images, called FM-VOC Dataset18644, containing depictions of structure fires, indoor fires, forest fires, and more was utilized to retrain the YOLOv5 model. The images were also manipulated — by flipping and rotating them, and also adjusting the brightness level — to make the model more robust where the scene is complex or lighting levels are challenging.

Furthermore, a FocalNextBlock module from a CFnet network was added to the system architecture. This assists the model in comprehending more complex scenes, and also reduces the number of model parameters, which slices the amount of computing power that is needed. An FPN network (QAHARep-FPN) was also integrated into the design of the system to reduce the number of redundant calculations that must be performed, further enhancing inference efficiency. A few other tweaks made by the team served to enhance the algorithm’s overall level of accuracy.

While this work undoubtedly moves the field forward, it may still be a while before this sort of technology is ready for the real world. It was noted, for example, that the system is capable of detecting smoke with greater precision than fire. The researchers believe that this is due to the greater complexity of features associated with fires. In any case, this issue will need to be addressed before a larger-scale rollout of the detector can be considered.