Gas Stove Flame Detection – Detect if gas knob is on or not. 
Team Members
- Ripinjal Rahi
- Parth Dhumak
- Supriyo Rana
- Satyam Tiwari
Use Case Importance
Detecting whether a gas stove knob is on or off is crucial for preventing potential fire hazards, gas leaks, and energy waste in households. Early detection can enhance home safety by alerting users to unattended or malfunctioning stoves, reducing the risk of accidents.
Data Collection and Annotation
We collected approximately 100+ real-world images of gas stoves from our house, college canteen and nearby hotels in different lighting.
We used ROBOFLOW and CVAT for annotation and labelling.
Classes annotated : “knob_ON”, “knob_OFF”
No. of classes = 2
Images were labeled and split into training, validation, and test sets. Data was augmented using brightness adjustments and flipping for better generalization.
train - 134 images
valid - 30 images
test - 21 images
Sample Images:
Model Training and Validation
We trained a YOLOv9s model for the gas stove flame detection task due to its enhanced accuracy and efficiency, making it ideal for real-time safety monitoring applications.
Key performance metrics monitored during training included mAP, precision, and recall to ensure reliable detection of ‘knob on’ and ‘knob off’ states.
To improve accuracy, we increased the number of training epochs and the input image size, allowing the model to learn more detailed features.We collected additional data to improve model performance to reduce misclassification and enhance accuracy.
Model Deployment and Demo Video
In real-world testing, the YOLOv9s model achieved an average accuracy of 92% - 98% in detecting the gas stove flame state, with a processing speed of around 25–30 FPS on compatible devices.
While the model runs smoothly on iOS devices, we observed occasional lag on some Android devices, likely due to hardware limitations or differences in GPU acceleration.
Demo Video
Watch the model in action: Gas Stove Flame Detection 
Real-time flame detection with YOLOv9s on mobile devices.
Conclusion
We learned how to train and deploy a real-time object detection model using YOLOv9s, evaluate its performance using key metrics like mAP and precision, and optimize it for mobile applications. We also gained practical experience in data collection, model tuning, and handling cross-platform performance issues.
