"Worker safety isn’t just a protocol—it’s the foundation of every industrial operation. But what if we could go beyond traditional checklists and PPE compliance? What if safety could be seen, understood, and acted upon in real time? What are the current systems available and what software, tools and algorithms are best to work with."
In this blog, we explore how deep learning, real-time tracking, and risk modeling can transform traditional PPE compliance systems into dynamic, predictive safety frameworks. We’ll walk through our proposed architecture, highlight key tools and models, and most importantly, draw insights from cutting-edge research that shaped our solution.
Literature Survey: Evolving Frontiers in AI-Driven Industrial Safety
Industrial safety has long been a critical domain of research, and its evolution has accelerated with the integration of computer vision and machine learning. Our literature review draws insights from multiple IEEE publications, focusing on the tools, algorithms, and models that have shaped the current landscape of AI-powered safety systems.
Here are five standout studies that informed our design:
[1] UAV-Based Crowd Monitoring
A real-time system using drone-mounted cameras successfully detected individuals and triggered alerts based on risk zones. Alerts were tiered—IT/Telegram for low-risk, voice alarms for high-risk—and the system operated efficiently on low-cost hardware (2–17 FPS, 65–80% CPU usage). Future enhancements include predictive movement analysis and behavior classification.
[2] PPE Usage Classification with BiLSTM
Among various tested architectures, BiLSTM outperformed others in recognizing PPE usage. Engineered features had minimal impact, and the study introduced a novel, publicly available dataset focused on PPE compliance.
[3] Real-Time Detection with Advanced Models
Models like CenterNet, Vision Transformer, and YOLOv7 were used for high-speed PPE detection. Leveraging large CCTV datasets and inference optimization techniques, the system achieved improved accuracy and responsiveness for workplace safety.
[4] Dataset Scale and Real-World Complexity
This study addressed limitations of earlier systems that relied on synthetic or narrow datasets. By using authentic surveillance footage and evaluating multiple architectures, the authors improved detection precision and extended compliance monitoring to complex industrial scenarios with variable lighting, occlusions, and crowd density.
[5] Edge AI for Privacy-Preserving Monitoring
Edge-based PPE detection using YOLOv4-Tiny enabled fast, reliable monitoring without internet dependency. The model balanced speed and accuracy on embedded hardware, and diverse datasets enhanced robustness across real-world conditions.
System Architecture: From Detection to Prediction
Our solution uses a modular pipeline integrating detection, tracking, and risk modeling:
PPE Detection: YOLOv8 detects helmets, vests, gloves, boots, and masks per video frame, outputting bounding boxes and PPE confidence scores.
Object Tracking & Motion Analysis: DeepSORT assigns unique IDs to workers, tracking position, velocity, acceleration, and proximity to hazards.
Risk Scoring Algorithm (RSA):
Computes a safety index using hazard distance, velocity, acceleration, PPE compliance, and crowd density, normalized via sigmoid.
- Risk Classification:
Scores are categorized as Safe (<0.4), Caution (0.4–0.7), or High Risk (≥0.7), with a temporal filter requiring 5 consecutive high-risk frames.
Predictive Risk Modeling (Optional): BiLSTM forecasts future risk scores for proactive safety measures.
Integration Workflow: Real-time video capture → PPE detection → tracking → risk scoring → classification → alerts via dashboard or audio.
Research Gaps & Future Directions
Despite progress, key challenges remain:
Data Fusion & Edge Constraints – Multi-camera integration and real-time edge processing are still limited.
PPE Ambiguity – Models struggle to distinguish between wearing vs. carrying gear.
Dataset Diversity – Many datasets lack scale and environmental variation.
Audit Integration – Process mining is underutilized in safety audits.
Lighting & Occlusion – Detection fails under poor visibility; multi-sensor fusion is underexplored.
Privacy & Ethics – Surveillance systems need stronger safeguards.
Conclusion: Smarter, Safer, Scalable
Our framework moves beyond traditional PPE detection by integrating motion dynamics, hazard proximity, and predictive modeling. It’s scalable, real-time, and adaptable to complex industrial environments.
By learning from existing research and addressing its gaps, we’re building a smarter safety system—one that doesn’t just react but anticipates.
References:
[1]A. A. Ardebili, M. Zappatore, A. Longo, and A. Ficarella, “Virtual Fencing for Safety-Critical Cyber-Physical Systems: Computer-Vision Enabled Digital Twins,” IEEE Access, vol. 13, pp. 12–31, 2025.
[2]P.C. da Fonseca Guimarães, L. B. de Cristo, M. E. P. Monteiro, J. L. Rebelatto, G. de S. Peron, O. K. Rayel, and G. L. Moritz, “Deep Learning for Real-Time PPE Usage Monitoring Using Wearable IMU Sensors,” IEEE Journal on Selected Topics in Signal Processing, vol. 13, no. 2, pp. 492–501, 2025.
[3]S. Al-Azani, H. Luqman, M. Alfarraj, A. A. I. Sidig, A. H. Khan, and D. Al- Hamed, “Real-Time Monitoring of Personal Protective Equipment Compliance in Surveillance Cameras,” IEEE Access, vol. 12, pp. 882–893, 2024.
[4]M. Imran, S. Hamid, and M. A. Ismail, “Advancing Process Audits with Process Mining: A Systematic Review of Trends, Challenges, and Opportunities,” IEEE Transactions on Engineering Management, vol. 11, no. 4, pp. 212–231, 2023.
[5]G. Gallo, F. Di Rienzo, F. Garzelli, P. Ducange, and C. Vallati, “A Smart System for Personal Protective Equipment Detection in Industrial Environments Based on Deep Learning at the Edge,” IEEE Access, vol. 10, pp. 62–78, 2022.
This work is done under the guidance of Ms.D.Diana Josephine, Associate Professor, Coimbatore Institute of Technology and along with my teammates Khushi Tirkey, Krisha S V, Nivedha M, and finally myself Sriyaa G B
SOCIAL SHARE CARD GENERATOR