Using Deep Learning with Thermal Imaging for Human Detection in Heavy Smoke Scenarios.

Area of Science:

• Fire Safety Engineering
• Artificial Intelligence
• Computer Vision

Background:

• Emergency evacuations in smoky fire scenarios pose significant risks due to low visibility.
• Traditional detection methods are often ineffective in these challenging environments.
• Advanced sensing and AI are needed for real-time situational awareness.

Purpose of the Study:

• To develop and evaluate an intelligent human detection system for low-visibility smoky fire evacuations.
• To leverage thermal imaging and deep learning for accurate and real-time people localization.
• To enhance firefighter safety and improve rescue response times.

Main Methods:

• Utilized a thermal imaging camera (TIC) capturing low-wavelength infrared (LWIR) images compliant with National Fire Protection Association (NFPA) 1801 standards.
• Employed the YOLOv4 deep learning model for real-time object detection on the acquired thermal images.
• Trained the YOLOv4 model on a single Nvidia GeForce 2070 GPU.

Main Results:

• The YOLOv4 model achieved over 95% precision in detecting people's locations in low-visibility smoky conditions.
• The system demonstrated real-time performance with a processing speed of 30.1 frames per second (FPS).
• LWIR thermal imaging proved effective for human detection through smoke.

Conclusions:

• The proposed thermal imaging camera and deep learning approach offer a viable solution for intelligent human detection during fire evacuations.
• Real-time detection capabilities provide critical information for control centers, enabling timely rescue and enhancing firefighter safety.
• This technology can significantly improve safety protocols in hazardous smoky environments.