Video Smoke Detection Method Based on Change-Cumulative Image and Fusion Deep Network

Area of Science:

• Computer Vision
• Fire Detection
• Deep Learning

Background:

• Traditional smoke detection relies on color, shape, texture, and motion, but these features lack robustness in complex environments.
• Existing deep learning methods struggle with detailed feature extraction and effective utilization of smoke motion characteristics, leading to high false alarm rates.

Purpose of the Study:

• To enhance smoke detection performance in videos, particularly under challenging environmental conditions.
• To develop a novel method that effectively captures both motion and color-change characteristics of smoke.

Main Methods:

• A change-cumulative image is generated by converting multi-frame video images from YUV color space.
• A fusion deep network is designed, enhancing VGG16 and Resnet50 (Deep residual network) by increasing network depth for improved feature expression.
• The change-cumulative image serves as input to the fusion deep network.

Main Results:

• Using change-cumulative images as input significantly outperforms traditional RGB input for smoke detection.
• The proposed fusion deep network demonstrates superior performance compared to standalone VGG16 and Resnet50 models.
• The method achieves better accuracy, lower false positive rates, and reduced false alarm rates than current popular video smoke detection techniques.

Conclusions:

• The change-cumulative image combined with a fusion deep network offers a robust and effective solution for video smoke detection.
• This approach significantly improves detection accuracy and reduces false alarms, making it suitable for real-world applications.
• The method advances the field of computer vision for fire detection systems.