Fire Detection using Image Processing

This repository presents a hybrid, vision-based fire detection system built using YOLOv8, classical image-processing techniques, and temporal analysis for flicker and motion. The project demonstrates how modern deep learning can be combined with lightweight computer vision modules to achieve real-time and accurate indoor fire detection.

---

Project Overview

The system takes live or recorded video input and detects fire and smoke using a hybrid framework that fuses:

• YOLOv8 for deep-learning–based fire/smoke object detection
• HSV color segmentation to isolate fire-colored pixels (orange–yellow hues)
• Motion detection to ensure detected regions are dynamic
• Flicker analysis to validate temporal irregularities typical of flames
• Circularity filtering to reject bright, static objects (e.g., bulbs, sunlight reflections)

Once a fire event is confirmed across multiple frames, a threaded alarm is triggered to provide an instant alert while ensuring smooth video playback.

---

Repository Structure

 Fire Detection Repository
 ┣  Fire-Detection
 ┃ ┗ Code implementation and detection logic (main hybrid model)
 ┣  Sample_videos
 ┃ ┗ Example videos to test the fire detection system
 ┣  UNISA_Dataset
 ┃ ┗ Partial dataset used for model evaluation (due to size constraints)
 ┣  Dataset_evaluation.py
 ┃ ┗ Python script to evaluate model performance on the dataset
 ┣  model.py
 ┃ ┗ Main program file integrating YOLOv8 and hybrid confirmation logic
 ┣  README.md
 ┗ (Other supporting files)

---

Usage

1. Clone this repository:

   git clone https://github.com/Akshat9936/Fire-Detection.git
   cd Fire-Detection

2. Run the main model:

   python model.py

3. To test with your own videos:

   • Place the video inside the Sample_videos/ folder.
   • Update the VIDEO_PATH variable in the code to point to your file.

4. To evaluate the model on the UNISA dataset:

   python Dataset_evaluation.py

---

UNISA Fire Detection Dataset

The UNISA (MIVIA/UNISA Fire Detection Dataset) is a widely used benchmark designed for evaluating vision-based fire detection systems. It contains 31 videos, including:

• 14 confirmed fire sequences recorded under different indoor conditions, and
• 17 non-fire videos (e.g., reflections, colored lights, moving red/orange objects) specifically created to challenge false-alarm handling.

This dataset allows performance assessment in terms of true positives, false positives, and overall detection accuracy.

Note: Due to GitHub’s upload size limitations, only a subset of the dataset is provided here for reference. For complete testing and reproducibility, it is strongly recommended to download the full dataset from its original source below:

Original UNISA Fire Detection Dataset: https://mivia.unisa.it/datasets/video-analysis-datasets/fire-detection-dataset/

---

Evaluation Script

The Dataset_evaluation.py program can be used to evaluate detection performance on the dataset. It compares predicted bounding boxes and temporal detections across frames to compute:

• True Positive Rate (TPR)
• False Alarm Rate (FAR)
• Frame-wise accuracy
• Detection latency per sequence

This ensures objective benchmarking of the proposed model against the dataset’s challenging conditions.

---

Requirements

Install dependencies before running:

pip install opencv-python numpy ultralytics torch

Optional (for Windows alarm functionality):

pip install winsound

---

Key Highlights

• Real-time inference: ~24 FPS on standard CPU
• High accuracy: ~90% overall detection accuracy on mixed indoor test videos
• Low false alarm rate through hybrid fusion logic
• Modular design for easy customization and dataset evaluation

---

Conclusion

This project demonstrates an efficient and practical approach to real-time indoor fire detection using video analytics. It integrates the reliability of deep learning with explainable classical methods, achieving robust, fast, and cost-effective fire safety automation.

---