-
Use Case Title
Helmet vs No Helmet (Two-Wheelers) - Detect compliance of safety gear.
-
Team Members
1.Mayur Haldankar
2.Sahil Dongre
3.Ankitkumar Gaud
4.Pareen Churi
5.Dhruvesh Ghadiali -
Use Case Important
Ensuring that workers wear helmets is crucial for preventing head injuries in hazardous environments like construction sites and factories. Automated detection systems can provide real-time monitoring, reducing the reliance on manual checks and enhancing overall safety compliance.
-
Data Collection and Annotation
-
Data Collection: Short video clips were recorded on public roads
using a smartphone camera. These recordings captured a variety of real- life scenarios including people riding bikes with and without helmets. The videos were taken under different lighting conditions (daylight,overcast, shadows) and from varied angles to simulate real-world complexity and improve generalization. -
Annotated Classes: Double class: Helmet and No Helmet.
-
Annotation Tool: Roboflow.
-
Total Images: 3000 (2100 train, 600 val, 300 test).
-
Sample Images:
-
-
Model Training and Validation
1.We trained a YOLOv9 model using a dataset of 2,100 training images and
600 validation images, specifically collected from road scenarios involving
motorbike riders wearing and not wearing helmets.-
Metric Monitors: Precision, Recall, mAP, and Loss
-
During early training runs, the model achieved high precision for detecting “Helmet” but showed false positives for “No Helmet” in busy area.
-
-
Model Deployment and Demo Video
-
The trained YOLOv9 model (best.pt) was successfully deployed to the YOLOvX mobile app for real-time helmet compliance monitoring.
-
The model was tested directly on mobile devices, showcasing its ability to detect “Helmet” and “No Helmet” classes accurately in real-world conditions such as roads with varying lighting, motion, and background clutter.
-
Inference Speed: Achieved approximately 5–10 FPS on mid-range smartphones
-
Limitations: Minor latency and occasional frame drops were noted on lower-end devices, especially during complex scenes with multiple subjects or fast motion
-
A real-time video demo was recorded, showing the model actively detecting helmet usage in outdoor traffic settings. The detection was consistent across different angles, vehicle speeds, and rider distances from the camera
6.Demo link Video: https:https://drive.google.com/file/d/1qqGuSPNUDfBKGaYMGYP_OrMeMi8OrRFe/view?usp=sharing
-
-
Conclusion
Our helmet compliance detection model, trained using YOLOv9, achieved an accuracy of over 80% during real-world testing on roads and traffic environments. The system consistently identified riders with and without helmets, even under varied lighting, motion, and environmental conditions.
