Hard-Hat Detection with YOLOv8

Custom-trained PPE (hard hat) detection for industrial safety monitoring. 97.5% mAP@50 on a held-out set built from real factory CCTV footage.

This is the detection model behind sergak-ai, a real-time workplace safety monitoring system.

---

Why this exists

Off-the-shelf hard-hat models are trained on clean, well-lit stock photography and fall apart on real CCTV: low resolution, motion blur, backlit doorways, workers at 20+ metres, hats in a dozen colours under sodium lighting.

This repository documents the dataset work and training setup that produced a model that actually holds up on footage from Uzbek industrial sites.

---

Results

Model: yolov8n · 640px · batch 16 · 2 classes (helmet, no_helmet)

Metric	Value
mAP@50	0.975
mAP@50-95	0.822
Precision	0.956
Recall	0.928

Run comparison

Run	Epochs	Precision	Recall	mAP@50	mAP@50-95
helmet_exp_gpu	10	0.956	0.928	0.975	0.822
helmet_exp55	5	0.919	0.898	0.957	0.794
train10	5 (early stop)	0.828	0.854	0.906	0.692

The jump from run to run came almost entirely from dataset work, not hyperparameters.

Nothing in these tables is typed by hand. Every figure comes from the Ultralytics output committed in reports/ — the per-epoch history is in reports/results.csv and the exact hyperparameters in reports/args.yaml.

---

Dataset

Classes	helmet, no_helmet
Sources	Public hard-hat datasets, merged and re-labelled, plus frames pulled from the deployment cameras
Labelling	Self-hosted CVAT (Docker), exported in YOLO format
Merge pipeline	scripts/ — download, deduplicate, remap class ids, split
Augmentation	Mosaic, HSV jitter, random scale, horizontal flip

What actually moved the number: adding frames sampled from the cameras the model would run on. Domain match beat both dataset size and every hyperparameter sweep. A model at 90% on stock photos dropped to the low 70s on site footage until real frames went into training.

Dataset config: data/helmet.yaml. The images themselves are not committed — point path at your own copy.

Pipeline scripts

Script	Does
scripts/download_roboflow_datasets.py, download_kaggle.py, download_huggingface.py, download_github.py	pull source datasets
scripts/convert_voc_to_yolo.py	VOC XML annotations → YOLO txt
scripts/prepare_sh17.py, prepare_new_datasets.py, prepare_more_datasets.py	normalise each source to the two-class scheme
scripts/merge_datasets.py, final_merge.py	deduplicate and merge into one train/val/test split
scripts/validate_dataset.py, check_datasets.py, stats.py	label sanity checks and class balance
scripts/visualize.py	draw labels over images for spot-checking
scripts/test_inference.py, live_test.py, test_gui.py	single-image, RTSP and GUI testing
scripts/export_production.py	export the chosen run's best.pt for deployment

---

Reproduce

pip install -r requirements.txt

# train
python train.py --data data/helmet.yaml --model yolov8n.pt \
                --epochs 10 --imgsz 640 --batch 16

# evaluate
python val.py --weights weights/best.pt --data data/helmet.yaml --split test

# run on an image, a video, or a live stream
python predict.py --weights weights/best.pt --source path/to/image.jpg
python predict.py --weights weights/best.pt \
                  --source "rtsp://user:pass@192.168.1.64:554/Streaming/Channels/101"

Training configuration

These are the actual arguments of the run that produced the numbers above, copied from reports/args.yaml:

model:      yolov8n.pt
epochs:     10
imgsz:      640
batch:      16
optimizer:  auto
lr0:        0.01
patience:   100
device:     cpu
seed:       0

Two things worth flagging, because the file is in the repository and you will see them:

• The run is named helmet_exp_gpu but device: cpu — the name is left over from an earlier attempt and the winning run happened to finish on CPU. Ten epochs of yolov8n at 640px is cheap enough that it did not matter.
• train.py ships with heavier defaults (yolov8l, 150 epochs, AdamW, batch 6) aimed at an 8 GB GPU. The shipped model is the lighter yolov8n run above, chosen for edge inference speed rather than peak accuracy. Pass the arguments shown here to reproduce it.

---

Repository layout

data/helmet.yaml     dataset config (2 classes, split paths)
scripts/             dataset download, convert, merge, validate, export
train.py             training entrypoint
val.py               evaluation — prints precision, recall, mAP per class
predict.py           inference on image / video / folder / RTSP
reports/             metrics, curves, confusion matrix, sample predictions
requirements.txt     pinned dependencies
weights/             released model weights (see Releases)

Weights are distributed through Releases rather than committed, to keep the repository small.

---

Limitations

• Trained on 2 classes only — it does not distinguish hat colour or detect other PPE (vests, goggles, gloves)
• Performance degrades below roughly 100px of person height in frame
• Night / IR footage was under-represented in training; expect lower recall there
• yolov8n was chosen for edge deployment speed; a larger backbone would likely gain a few points of mAP@50-95
• The reported split is the one shipped with the merged dataset; it is not a public benchmark, so these numbers are not directly comparable to published hard-hat results

---

License

MIT — see LICENSE. Weights are released for research and evaluation use.

Author

Mardonbek Sulaymonqulov — AI / Computer Vision Engineer GitHub · mardonbeksulaymonqulov156@gmail.com