Workplace safety vision system with privacy-aware outputs
This project builds a visual monitoring system for construction-site safety. It detects workers, safety equipment, and violations, supports image/video/webcam inputs, generates reports, and includes a privacy mode that blurs faces automatically.
Challenge
- Construction safety monitoring needs fast detection of missing protective equipment.
- Visual systems should support multiple input modes and practical review workflows.
- Monitoring people raises privacy concerns, so the interface needs built-in privacy protection.
System architecture
Data and inputs
- Roboflow Site Construction Safety Dataset.
- Started with 2,777 images and expanded to 6,630 images through augmentation.
- The isolated test set contains 287 images resized for the detection workflow.
Technical approach
- Train a YOLOv8-based detection model for PPE and construction-site objects.
- Detect classes such as Hardhat, Safety Vest, Mask, Person, Machinery, Vehicle, and violation classes such as NO-Hardhat.
- Build a Streamlit interface for image upload, clipboard input, video processing, and webcam snapshot analysis.
- Add class filtering, confidence threshold control, CSV reporting, and face blurring.
Evaluation and results
10 object classes
6,630 augmented images
mAP@50 80.2% · precision 89.3%
- The system reports mAP@50 of 80.2%, precision of 89.3%, and recall of 71.7% on the test setup.
- The interface supports image, video, and webcam-based demonstrations.
- Privacy mode automatically blurs detected faces while preserving safety-compliance context.
- Class filtering and CSV export make the results easier to review after analysis.
Implementation and code
Implementation focus
The implementation connects data preparation, modeling, evaluation, and interpretation in a structured workflow that makes the technical decisions clear.
Source code
The code is available for exploring the implementation details and extending the experiment when needed.
Scope and responsible use
The project is a focused modeling and evaluation study. Broader use should be supported by validation on additional data, robustness checks, monitoring, and domain-specific evaluation.
Future development
- Improve recall for smaller or visually ambiguous objects.
- Add stronger field validation across sites, cameras, and lighting conditions.
- Extend reporting with trend dashboards and threshold calibration.
Technical contribution
The project demonstrates applied visual AI for operational safety: combining detection, interface design, privacy protection, and reporting into one practical workflow.