Real-time Personal Protective Equipment (PPE) detection for construction and industrial environments. Trained on a merged dataset of 9,675 images using YOLOv5s and YOLOv5m architectures.

Published at AAAI 2025 Summer Symposium → Read the Paper

📁 GitHub Repository → Darth-Freljord/ppe-detection-yolov5

🗂️ Model Files

File	Description	mAP@0.5	FPS
`yolov5m_a_best.pt`	Best model — YOLOv5m with anchor evolution, 896×896 input	0.841	25–30
`yolov5s_a_best.pt`	Lightweight — YOLOv5s, 640×640 input	0.736	35–40
`yolov5s_b_best.pt`	Transfer learning variant — YOLOv5s	0.683	35–40

⚡ Quick Start

python

import torch

# Load best model
model = torch.hub.load('ultralytics/yolov5', 'custom',
                        path='yolov5m_a_best.pt')

# Run inference on an image
results = model('construction_site.jpg')
results.show()
results.print()

Download with huggingface_hub

python

from huggingface_hub import hf_hub_download

model_path = hf_hub_download(
    repo_id="Darth-Freljord/ppe-detection-yolov5",
    filename="yolov5m_a_best.pt"
)

import torch
model = torch.hub.load('ultralytics/yolov5', 'custom', path=model_path)

📊 Performance

Overall Results

Model	mAP@0.5	mAP@0.5:0.95	Precision	Recall	FPS
YOLOv5m(a) (recommended)	0.841	0.649	0.864	0.776	25–30
YOLOv5s(a)	0.736	0.409	0.748	0.721	35–40
YOLOv5s(b)	0.683	0.360	0.682	0.662	35–40

YOLOv5m(a) — Per-Class Performance

Class	Precision	Recall	mAP@0.5	mAP@0.5:0.95
Mask	0.948	0.945	0.969	0.850
Helmet	0.873	0.833	0.901	0.713
Vest	0.852	0.762	0.818	0.640
Glasses	0.860	0.744	0.812	0.538
Person	0.860	0.816	0.890	0.674
Gloves	0.790	0.555	0.655	0.480

🏷️ Classes Detected

0: Person
1: Helmet
2: Vest
3: Gloves
4: Glasses
5: Mask

No-PPE Detection: The model also dynamically detects missing PPE (e.g. No-Helmet, No-Vest) at inference time without requiring additional trained classes. See detect_mod.py in the GitHub repo.

📁 Training Data

Three datasets were merged into a unified dataset of 9,675 images (70:15:15 split):

Dataset	Images	Classes	Annotation Format
Pictor-PPE (ciber-lab, 2019)	784	Person, Helmet, Vest	Tab-separated
VOC2028 (njvisionpower, 2023)	7,581	Person, Helmet	PASCAL VOC XML
CHV (ZijianWang-ZW, 2020)	1,330	Person, Vest, Helmet (multi-color)	YOLO

The extended dataset used for YOLOv5m(a) was further augmented using a pretrained YOLOv9 model to generate additional Mask, Gloves, and Glasses labels.

⚙️ Training Configuration (YOLOv5m(a))

Parameter	Value
Input Resolution	896×896
Batch Size	4
Epochs	30 (initial) + 100 (post-evolution)
Frozen Layers	8 (initial), 0 (retraining)
Anchor Evolution	100 iterations
Learning Rate (lr0)	0.005
Momentum	0.937
Mosaic Augmentation	0.768
Base Weights	`yolov5m.pt` (COCO pretrained)

📄 Citation

If you use this model, please cite:

bibtex

@inproceedings{ahmed2025seeing,
  title={Seeing Safety: Computer Vision for Real-Time PPE Monitoring in the Middle East Construction Sector},
  author={Syed Suhail Ahmed},
  booktitle={AAAI 2025 Summer Symposium: Context-Awareness in Cyber Physical Systems},
  year={2025},
  url={https://ojs.aaai.org/index.php/AAAI-SS/article/view/36044}
}

DarthRegicid1/YOLOv5_PPE-Detection

🦺 PPE Detection — YOLOv5 (Real-Time)