Transformer-Based Fire Detection in Videos.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Effective fire detection in surveillance is crucial for preventing hazardous situations.
• Accurate and fast models are essential for real-time video analysis.
• Existing methods may lack the speed or precision required for immediate threat response.

Purpose of the Study:

• To propose a novel transformer-based network for accurate and real-time fire detection in videos.
• To evaluate the model's performance on both full-frame fire classification and precise fire localization tasks.
• To demonstrate the model's superiority over state-of-the-art methods.

Main Methods:

• An encoder-decoder transformer architecture was developed for video frame analysis.
• The model computes attention scores to identify relevant regions for fire detection.
• The methodology was trained and evaluated for fire/no-fire classification and pixel-level fire localization.

Main Results:

• The proposed model achieved 97% accuracy in full-frame classification.
• Real-time processing at 20.4 frames per second (fps) was demonstrated.
• Excellent fire localization performance with a 0.02 false positive rate and 97% f-score/recall was achieved.

Conclusions:

• The transformer-based network offers a highly accurate and efficient solution for real-time video fire detection.
• The model excels in both identifying the presence of fire and pinpointing its exact location within video frames.
• This approach represents a significant advancement for surveillance systems requiring rapid and precise fire identification.