Fire incidents can cause catastrophic damage and loss of life when not detected early enough. Current detection methods, which rely primarily on traditional sensors, often have delayed response times that can lead to severe consequences. This creates a critical need for faster, more reliable detection systems that can identify fire hazards at their earliest stages. To address this challenge, this project implements advanced deep learning technology to detect fires and smoke as they emerge, enabling rapid response and intervention.
Flare Guard represents a state-of-the-art approach to fire safety monitoring. At its core, the system uses YOLOv11, a powerful object detection model, to continuously analyze video feeds for signs of fire and smoke in real-time. When potential threats are detected, the system immediately sends alerts through Telegram/WhatsApp messaging platforms, ensuring that stakeholders are notified without delay.
Key capabilities:
- Real-time video processing with 30-60 frames per second
- High-accuracy threat identification (80.6% Precision)
- Robust performance across various environmental conditions
- Immediate alert system with visual confirmation via Telegram and WhatsApp.
- Adaptable for multiple use cases including surveillance systems, industrial monitoring, home security, and robotic applications
The combination of speed, accuracy, and reliable notification makes Flare Guard particularly effective for environments where early fire detection is critical for safety and asset protection.
Instant_Notification_System.mp4
This video demonstrates the system detecting fire immediately upon occurrence.
Alerts are sent instantly to Telegram and WhatsApp, including an attached image.
Thanks to multithreading, the system continues running without interruption.
The dataset consists of 10,463 annotated images, available on Roboflow. This dataset is designed for training and evaluating object detection models tailored for real-time fire and smoke detection. It is suitable for:
- Surveillance systems (CCTV monitoring, smart security cameras)
- Industrial safety applications (factories, warehouses, refineries)
- Residential safety solutions (smart home fire detection)
- Autonomous monitoring systems (drones, robotics, IoT devices)
| Split | Images | Annotations |
|---|---|---|
| Training | 9,156 | 27,468 |
| Validation | 872 | 2,616 |
| Test | 435 | 1,305 |
Classes: Fire, Smoke
Annotation Format: YOLOv11-compatible bounding boxes
# Download dataset via Roboflow
from roboflow import Roboflow
rf = Roboflow(api_key="YOUR_API_KEY")
project = rf.workspace("sayed-gamall").project("fire-smoke-detection-yolov11")
dataset = project.version(2).download("yolov11")The model was trained using YOLOv11 on a dataset of fire and smoke images. Training stopped early due to no improvement over 20 epochs, with the best results observed at Epoch 92.
| Metric | Value |
|---|---|
| Precision (P) | 0.806 |
| Recall (R) | 0.717 |
| mAP@50 | 0.770 |
| mAP@50-95 | 0.492 |
| Class | Precision | Recall | mAP@50 | mAP@50-95 |
|---|---|---|---|---|
| Fire | 0.813 | 0.806 | 0.828 | 0.513 |
| Smoke | 0.800 | 0.629 | 0.711 | 0.472 |
Here are examples from the test set:
-
Clone the repository:
git clone https://github.com/sayedgamal99/Real-Time-Smoke-Fire-Detection-YOLO11 cd Real-Time-Smoke-Fire-Detection-YOLO11 -
Install the required packages:
pip install ultralytics
To perform inference with the trained model on test images, run:
yolo detect predict model=models/best_nano_111.pt source=data/house.png conf=0.35 iou=0.1and there's the output:
To perform inference in real-time using a webcam:
yolo detect predict model=models/best_nano_111.pt source=0 conf=0.35 iou=0.1 show=TrueProtect What Matters Most - Early Detection Saves Lives