Client Overview
A growing eBike fleet operator in India faced a recurring issue: riders not wearing helmets led to automatic fines issued to the fleet owner, not the individual rider. With increasing fines and no way to verify accountability, the client needed a smart, real-time detection solution to protect their business.
The Challenge
In India, roadside surveillance cameras enforce helmet laws. Violations result in e-tickets sent to vehicle owners. For fleet managers, this meant:
- Mounting liability for rider behavior.
- Financial penalties with no clear recourse.
- A need for evidence-based, automated compliance enforcement.
Our Solution
We developed a real-time Helmet Detection App using TensorFlow and Python, designed specifically for Android smartphones mounted on eBikes. The app detects if a rider is wearing a helmet and logs images accordingly.
Why Android?
- 70%+ market share in India.
- Cost-effective, globally available, IP-rated phones.
Data Collection & Model Training
- Volunteers provided videos with/without helmets.
- We used FFmpeg to convert videos into labeled image datasets.
- Dataset split: 70% for training, 30% for testing.
- Used back regression and iterative optimization to improve accuracy.
Model Deployment
- Integrated into the app to assess helmet usage every few seconds.
- Results were filtered using multiple validation layers:
- Image quality checks (noise, saturation, brightness).
- Face detection to confirm rider presence.
- Multi-frame consistency before logging a violation.
Scaling with Mechanical Turk
Initially, all image labeling was done manually. To scale up:
- We used Amazon Mechanical Turk to crowdsource accurate image labels.
- Trained the model iteratively using the expanding dataset.
Field Challenges & Fixes
Issue: Poor image quality due to sun glare, motion blur, and low-cost cameras.
Fixes:
- Discarding low-quality images using pre-screening filters.
- Taking multiple shots and accepting results only on consistent predictions.
- Embedding face recognition to ensure a valid detection zone.
Results
- Accurate helmet detection in real-world riding conditions.
- Fewer fines and better rider accountability.
- Real-time feedback loop to improve future predictions.
“Overall the project was a great success, and the App worked as expected.”
Technologies Used
- TensorFlow (ML framework)
- Python (Development language)
- Android OS (App platform)
- FFmpeg (Video to image conversion)
- Amazon Mechanical Turk (Image labeling at scale)
About Skillion
Pete Cooper is a CEO and Program Manager with over 20 years of experience across AI/ML, Product Design, Software, and Engineering. At Skillion, Pete and the team combine program management with cutting-edge tech to deliver results at scale.
Need help with an AI-powered product? 📩 Contact: info@skillion.tech
