A Lightweight Fire Detection Framework for Edge Visual Sensors Using Small-Sample Domain Adaptation.

Area of Science:

• Computer Vision
• Machine Learning
• Sensor Networks

Background:

• Vision-based sensor networks face challenges in accurate fire detection due to varying environmental conditions.
• Acquiring labeled data for diverse scenarios (day/night, interference) is difficult for training robust models.

Purpose of the Study:

• To develop a novel fire detection framework for vision-based sensor networks.
• To enhance fire detection accuracy and robustness across different scenarios using domain adaptation.

Main Methods:

• Constructing a high-dimensional feature vector by fusing HSI color statistics, Local Binary Pattern (LBP) textures, and Wavelet Transform shape features.
• Employing an Adaptive Support Vector Machine (A-SVM) with a small-sample domain adaptation mechanism to fine-tune models with limited target domain data.
• Utilizing regularization constraints for efficient model parameter adjustment.

Main Results:

• The proposed Multi-Feature Fusion and A-SVM method significantly improved F1-scores by 19% (daytime) and 30% (nighttime) compared to baseline methods.
• Demonstrated superior performance over traditional color thresholding and unadapted SVM classifiers in cross-domain scenarios.
• Achieved low-cost, high-precision, and robust fire detection.

Conclusions:

• The developed framework offers an effective solution for cross-scenario fire detection in vision-based sensor networks.
• The approach is well-suited for deployment on resource-constrained edge computing nodes.
• Domain adaptation is crucial for improving the performance of fire detection systems in diverse environmental conditions.