AI-Powered Industrial Safety: Real-Time Monitoring Using YOLOv8 and Edge AI
INTRODUCTION
In today’s rapidly advancing industrial landscape, worker safety is paramount. Industries such as construction, mining, oil refining, and manufacturing operate in environments filled with physical, chemical, and operational hazards. While strict safety regulations and Personal Protective Equipment (PPE) protocols exist, ensuring consistent compliance remains a persistent challenge. Human supervisors can overlook violations due to fatigue or limited visibility, and sensor-based systems often fail to provide a comprehensive view of workplace safety.
With the rise of Artificial Intelligence (AI) and Computer Vision, industries can now shift from manual oversight to automated, intelligent monitoring systems. Using deep learning models such as YOLO (You Only Look Once), AI can analyze video feeds in real time to detect whether workers are wearing helmets, gloves, and vests, or engaging in unsafe behaviors like smoking in restricted areas.
This blog presents a detailed overview of our project, “Real-Time Automated Safety Monitoring in Industrial Environments”, which employs YOLOv8 for real-time PPE and behavior detection on edge devices. It demonstrates how AI-powered safety monitoring can reduce accidents, improve compliance, and enhance workplace efficiency — all while operating autonomously and cost-effectively.
LITERATURE SURVEY ON EXISTING METHODS
Research into industrial safety monitoring has evolved dramatically, shifting from traditional supervision to vision-based deep learning techniques. A major focus has been the accurate detection of PPE items and unsafe behaviors under complex, dynamic industrial conditions.
1. Evolution of Safety Monitoring Systems
Earlier systems relied heavily on manual supervision, where safety officers visually confirmed PPE compliance. This method offered contextual understanding but suffered from human error, inconsistency, and lack of scalability. As industries grew in size and complexity, sensor-based monitoring emerged — using RFID tags, smoke detectors, and motion sensors to detect specific events. However, these systems lacked flexibility and often failed to distinguish between legitimate and unsafe conditions.
The emergence of computer vision and Convolutional Neural Networks (CNNs) enabled a paradigm shift. These models could automatically learn visual patterns, recognize safety equipment, and identify unsafe actions in real time — eliminating the need for handcrafted feature extraction.
2. YOLO-Based Research Advances
YOLO has been a breakthrough in real-time object detection due to its single-shot architecture that simultaneously predicts object locations and classes. Variants from YOLOv3 to YOLOv8 have progressively enhanced detection accuracy and speed, making them ideal for industrial applications.
MARA-YOLO (Di et al., 2024)
Di et al. introduced MARA-YOLO, an efficient model designed for multiclass PPE detection under complex industrial lighting and occlusion. It featured:
- A MobileOne-S0 re-parameterized backbone for lightweight computation.
- An Attentional Space-to-Depth (AS-Block) that preserved fine texture details.
- R-C2F and RASFF modules for adaptive multi-scale feature fusion.
Their custom KSE-PPE dataset achieved a 6.7% AP50 improvement over YOLOv8-s, proving MARA-YOLO’s superiority in PPE detection efficiency — especially for small or partially visible items.
Edge-Based YOLO (Gallo et al., 2022)
Gallo et al. explored edge computing for PPE detection, integrating YOLO models on low-power devices such as Raspberry Pi + Intel NCS2. The study compared YOLOv4, YOLO-tiny, SSD-MobileNetV2, and CenterNet models. Results showed YOLO-tiny provided the best real-time performance, confirming the practicality of deploying AI-based safety systems directly on embedded devices without relying on cloud servers.
YOLOv8n-ASF-DH (Lin, 2024)
Lin’s work enhanced helmet detection using advanced attention and fusion techniques:
- Triplet Attention Mechanism improved focus on small targets.
- Attentional Scale Sequence Fusion (ASF) optimized multi-scale detection.
- Dynamic Head (DyHead) allowed adaptive feature scaling.
This model achieved +2.58% mAP@0.5 compared to baseline YOLOv8n, demonstrating its robustness in detecting helmets in cluttered or dim environments.
3. Unsafe Behavior Detection Research
Detecting unsafe actions, such as smoking in restricted areas, remains challenging because such behaviors involve small, transient, and occluded objects. Two major YOLO-based approaches have addressed this:
YOLOv8-MNC (Wang et al., 2023)
Introduced a model optimized for smoking detection with Multi-Head Self-Attention (MHSA) and CARAFE upsampling. These modules captured global dependencies and reduced information loss, achieving 85.9% detection accuracy — a 5.7% improvement over baseline YOLOv8.
Smoking-YOLOv8 (Wang et al., 2024)
Enhanced detection for chemical plant safety by integrating:
- Selective and Dimension-Preserving Attention (SD-Attn)
- Wise-IoU (WIoU v3) loss for precision localization. This model achieved a 6.18% increase in mAP@0.5, excelling under variable lighting and high occlusion conditions.
4. Summary of Literature Insights
| Model | Focus | Techniques | Key Results |
|---|---|---|---|
| MARA-YOLO (2024) | PPE Detection | AS-Block, RASFF | +6.7% AP50 |
| Edge-YOLO (2022) | Edge Deployment | CNNs on Raspberry Pi | Real-time & Privacy-preserving |
| YOLOv8n-ASF-DH (2024) | Helmet Detection | Triplet Attention, ASF | +2.58% mAP@0.5 |
| YOLOv8-MNC (2023) | Smoking Detection | MHSA, CARAFE | +5.7% mAP@0.5 |
| Smoking-YOLOv8 (2024) | Behavior Monitoring | SD Attention, WIoU v3 | +6.18% mAP@0.5 |
These studies collectively emphasize that deep learning-based, edge-deployable, real-time systems outperform manual or sensor-based monitoring. They provide a robust foundation for developing scalable, AI-driven industrial safety solutions.
IDENTIFYING THE PROBLEM
Despite stringent safety regulations, industrial accidents continue to occur due to non-compliance with PPE and unsafe behavior. Manual supervision cannot guarantee consistent vigilance, and sensor-based systems are limited to detecting predefined conditions.
The key challenges identified are:
- Lack of Continuous Monitoring: Supervisors can’t observe every worker in large-scale operations.
- Inflexibility: Sensors can detect smoke or RFID tags but can’t confirm proper PPE usage.
- False Alarms: Traditional systems often trigger inaccurate alerts.
- Scalability: Expanding hardware-based systems to multi-site operations is costly and impractical.
To overcome these, there’s a pressing need for an automated, intelligent, and adaptive monitoring system that can detect multiple safety violations simultaneously and operate efficiently in real time.
DESIGNING THE COUNTERMEASURE: YOLOv8-BASED MONITORING
Our proposed approach leverages the YOLOv8 deep learning model for real-time detection of PPE compliance and unsafe behaviors. The system architecture integrates several key modules:
1. Input Module
Captures continuous video feeds from surveillance cameras installed in industrial zones.
2. Preprocessing
Frames are resized (640×640 or 416×416) and normalized. Data augmentation techniques — including rotations, brightness adjustment, and noise injection — enhance model generalization.
3. Detection Module
The YOLOv8 model identifies:
- PPE items (helmets, gloves, vests).
- Unsafe behaviors (smoking, absence of PPE).
The model uses Focal-EIoU loss and attention-based features for improved detection of small or occluded objects.
4. Decision Module
Analyzes YOLO outputs to determine compliance. Violations trigger alerts via:
- Audio alarms,
- SMS/email notifications, or
- Dashboard logs for audit purposes.
5. Edge Deployment
The system is optimized for devices like Raspberry Pi, Jetson Nano, and Intel NCS2, ensuring low latency and privacy-preserving, on-device inference.
IMPLEMENTATION DETAILS
Dataset Composition
Two datasets were developed:
- PPE Dataset: Images of workers with and without helmets, gloves, and vests.
- Unsafe Behavior Dataset: Smoking gestures, cigarettes, and lighter usage under varied conditions.
Model Training
- Framework: PyTorch + Ultralytics YOLOv8
- Optimizers: Adam / SGD with learning rate scheduling
- Loss Function: Focal-EIoU
- Split: 80% training, 10% validation, 10% testing
Training utilized early stopping and checkpointing to prevent overfitting, achieving strong convergence and high mean Average Precision (mAP) values.
Performance Metrics
- Helmet Detection: Precision 0.96, Recall 0.95
- Vest Detection: Precision 0.94, Recall 0.93
- Glove Detection: Precision 0.91, Recall 0.90
- Smoking Detection: Precision 0.92, Recall 0.91
- mAP@0.5: 0.905–0.955
These results validate the robustness of YOLOv8 in detecting diverse safety violations under challenging lighting and occlusion conditions.
RESULTS AND DISCUSSION
The system demonstrated reliable, real-time performance in both simulation and hardware deployment. On Jetson Nano, inference latency remained under 50 ms per frame, suitable for live CCTV applications.
Qualitative Results:
Visual tests confirmed accurate detection even in cluttered scenes with multiple workers. YOLOv8’s advanced attention modules improved the detection of small gloves and cigarettes — typically difficult targets.
Quantitative Analysis:
Precision–Recall curves showed balanced performance across PPE categories, with an average F1-score of 0.55 at optimal confidence thresholds.
Edge Deployment Benefits:
- Local processing ensures privacy.
- Minimal network dependency.
- Feasible for continuous 24/7 monitoring.
FUTURE WORK
Future development directions include:
- Expanding PPE detection to include boots, face shields, and ear protection.
- Integrating pose estimation to detect unsafe body postures or mobile phone usage.
- Enhancing robustness using thermal imaging for low-light monitoring.
- Employing federated learning to update models across sites without sharing raw data.
- Building a centralized dashboard for multi-camera analytics and automated compliance reporting.
These advancements will move the system toward a comprehensive, scalable industrial safety management platform.
CONCLUSION
The implementation of AI-powered industrial safety monitoring demonstrates how deep learning can revolutionize workplace safety. By leveraging YOLOv8 for PPE and behavior detection, industries can automate compliance enforcement, reduce accidents, and foster a proactive safety culture.
Unlike traditional systems, this solution:
- Operates in real time with high accuracy.
- Works effectively on low-power edge devices.
- Protects privacy through on-site processing.
This work highlights the intersection of AI, IoT, and edge computing in building a smarter, safer industrial future. As industries adopt such intelligent systems, the vision of zero-accident workplaces becomes increasingly attainable.
REFERENCES
- Di, W., Liu, X.F., & Zhou, Z. “MARA-YOLO: A Lightweight Multi-Class PPE Detection Model Based on YOLOv8.” IEEE Access, vol. 12, 2024.
- Gallo, S., Palmieri, F., & Castiglione, M. “Smart Edge Computing for Economic and Privacy-Preserving On-Device PPE Detection.” IEEE IoT Journal, 2022.
- Lin, C. “YOLOv8n-ASF-DH: Helmet Detection in Cluttered Industrial Environments.” IEEE Access, 2024.
- Wang, D., Xu, Y., & Huang, Z. “YOLOv8-MNC: Multi-Head Self-Attention Enhanced YOLO for Smoking Detection.” IEEE Sensors Journal, 2023.
- Wang, D., Chen, L., & Xu, Y. “Smoking-YOLOv8: High-Precision Unsafe Behavior Detection in Chemical Plants.” IEEE Transactions on Industrial Informatics, 2024.
Author Credits
This work was completed under the guidance of Dr. K. Harini, Assistant Professor, Coimbatore Institute of Technology, along with the project team: Adharsh C, Prem kumar S, Santhosh kumar G, and Selvakanthan B.
Top comments (0)