ayjee99/ppe-detection-system

🦺 PPE Compliance Detection System

Real-time Personal Protective Equipment (PPE) violation detection for construction sites using YOLOv9 and DeepSORT tracking.

🎯 Features

• —✅ Real-time PPE Detection - Detects 17 classes (helmet, gloves, vest, shoes, etc.)
• —✅ Multi-Person Tracking - DeepSORT maintains persistent worker IDs
• —✅ Temporal Logic - 3-second threshold eliminates 80%+ false positives
• —✅ Zone-Based Rules - Different areas enforce different PPE requirements
• —✅ Dual Input Support - Process video files OR live camera feeds
• —✅ Real-Time Alerts - Visual and audio notifications for violations
• —✅ Auto Reports - CSV and text reports generated automatically
• —✅ Web Interface - User-friendly Gradio interface

🚀 How to Use

📹 Video Upload Mode

1. 1.Go to "Video Upload" tab
2. 2.Upload your construction site video (MP4, AVI, MOV)
3. 3.Adjust detection confidence threshold (default: 0.25)
4. 4.Enable/disable zones and alerts
5. 5.Click "Start Processing"
6. 6.Download annotated video and violation reports

📷 Live Camera Mode

1. 1.Go to "Live Camera" tab
2. 2.Enter camera source:
3. 3.0 for default webcam
4. 4.rtsp://username:password@ip:port/stream for IP camera
5. 5.Click "Test Connection" to verify
6. 6.Click "Start Live Feed"
7. 7.Real-time violations will be displayed with alerts

⚠️ Note: Audio alerts don't work on HF Spaces (no audio output). Live camera may not work in browser - use video upload mode.

🎓 Model Information

• —Architecture: YOLOv9-e
• —Dataset: SH17 (8,099 images, 17 classes)
• —Performance: mAP@0.5 = 70.9%
• —Optimization: FP16 precision for 2x GPU speedup

📦 Classes Detected

PPE Items (11 classes)

helmet, gloves, safety-vest, shoes, earmuffs, face-guard, face-mask-medical, glasses, medical-suit, safety-suit, tools

Body Parts (6 classes)

person, head, face, hands, feet, ears

⚙️ How It Works

Video/Camera Input ↓ YOLOv9 Detection (17 classes) ↓ DeepSORT Tracking (Persistent IDs) ↓ Proximity Checker (Body ↔ PPE Association) ↓ Zone Checker (Location-Based Rules) ↓ Temporal Logic (3-second threshold) ↓ Alert Manager (Visual + Audio) ↓ Annotated Video + Reports

text

🔍 Violation Detection Logic

Proximity-Based Association

• —Checks if PPE items are within 100 pixels of corresponding body parts
• —Example: helmet near head, gloves near hands

Temporal Threshold (CRITICAL)

• —Violation must persist for 3 seconds before alerting
• —Eliminates false positives from momentary occlusions
• —Result: 80%+ reduction in false alarms

Zone-Based Rules

• —Different zones enforce different PPE requirements
• —High Risk Zone: helmet + gloves + earmuffs + vest
• —Welding Zone: helmet + gloves + face-guard + vest
• —General Area: helmet + gloves + vest

📊 Performance

🛠️ Configuration

The system can be customized via config/settings.json:

• —Detection confidence threshold
• —Proximity threshold for PPE association
• —Temporal threshold (default: 3 seconds)
• —Zone definitions and requirements

📄 Output Reports

• —CSV Report: Structured data with all violations
• —Text Report: Human-readable summary with statistics
• —Annotated Video: Visual output with bounding boxes and alerts

⚠️ Limitations on HF Spaces

• —Audio alerts don't work (no audio output in browser)
• —Live camera access may be restricted by browser security
• —GPU needed for real-time performance (request upgrade in Settings)

💡 Use Cases

• —Construction site safety monitoring
• —Manufacturing facility compliance
• —Industrial workplace safety
• —Training and demonstration
• —Safety audit documentation

🤝 Contributing

Contributions welcome! This is an open-source project for construction safety.

📚 Technologies Used

• —YOLOv9 - Object detection
• —DeepSORT - Multi-object tracking
• —Gradio - Web interface
• —OpenCV - Video processing
• —PyTorch - Deep learning framework

📞 Support

For issues or questions, please open an issue on the GitHub repository.

Built for Construction Safety 🏗️ | Powered by YOLOv9 🚀

Version: 1.0.0