Abstract:
In urban environments, fire detection is essential for public safety, as rapid
identification can prevent extensive damage and save lives. Traditional fire detection
methods rely on smoke sensors, which often suffer from delayed response times and
limited precision, particularly in complex environments. To address these limitations,
this project introduces an advanced fire detection system that combines traditional
image processing techniques with deep learning models, enhancing both accuracy and
speed in detecting fires. The system uses Candidate Region Detection, Feature
Extraction, and Classification stages, which incorporate multiple models and
techniques, including Gaussian Mixture Models, Support Vector Machines, and fine-
tuned deep learning models like ResNet50 and Inception V3.
Candidate regions likely to contain fire are identified based on color and texture
analysis. Fire-specific features are extracted using chromatic segmentation and texture
analysis with Local Binary Patterns and Gray Level Co-occurrence Matrices.
Additionally, flicker features are captured using Discrete Wavelet Transform to
distinguish fire from other dynamic objects. The classification phase uses Support
Vector Machines and fine-tuned CNNs to improve detection accuracy.
This system’s adaptability and modularity enable it to be integrated with existing
surveillance infrastructure, offering a robust solution for fire detection across various
urban settings. Extensive testing shows a high level of accuracy and a significant
reduction in false positives, making it suitable for real-world applications in residential,
commercial, and public spaces. Future developments may involve integrating
additional sensors and expanding the system's capabilities for more complex
environments
