---
license: agpl-3.0
library_name: ultralytics
pipeline_tag: object-detection
tags:
  - yolo
  - yolo11
  - yolo26
  - object-detection
  - worker-detection
  - person-detection
  - industrial
  - safety
  - computer-vision
base_model:
  - Ultralytics/YOLO11
---

# BalkonTech Models — Factory Worker Detection

Fine-tuned YOLO models for detecting **factory workers** in industrial environments. These models were trained on real-world factory footage to reliably localize workers under challenging conditions such as occlusion, machinery clutter, and varied lighting.

## Models

| File | Base model | Size | Task |
|---|---|---|---|
| `yolo11x_best.pt` | YOLO11x | 114 MB | Worker detection |
| `yolo26x_best.pt` | YOLO26x | 118 MB | Worker detection |

**Classes:** `worker` (person in a factory/industrial setting)

## Intended Use

- Worker presence detection on the factory floor
- Occupancy and zone-monitoring analytics
- Input stage for downstream safety systems (e.g., restricted-area alerts)

**Out of scope:** These models are not certified safety devices. Do not use them as the sole mechanism for life-critical decisions. Face recognition or identification of individuals is not supported and not intended.

## Usage

```python
from ultralytics import YOLO

# Load either model
model = YOLO("yolo11x_best.pt")   # or "yolo26x_best.pt"

# Inference on an image, video, or stream
results = model.predict("factory_frame.jpg", conf=0.4)

for r in results:
    for box in r.boxes:
        print(box.cls, box.conf, box.xyxy)
```

Download directly from the Hub:

```python
from huggingface_hub import hf_hub_download
from ultralytics import YOLO

weights = hf_hub_download(
    repo_id="etemkocaaslan/balkontech-models",
    filename="yolo11x_best.pt",
)
model = YOLO(weights)
```

## Training

- **Base models:** Ultralytics YOLO11x and YOLO26x pretrained weights
- **Data:** Proprietary dataset of factory-floor imagery annotated for workers
- **Fine-tuning:** Standard Ultralytics training pipeline

<!-- TODO: add training details (epochs, image size, augmentations, dataset size) -->

## Evaluation

<!-- TODO: fill in validation metrics -->

| Model | mAP50 | mAP50-95 | Precision | Recall |
|---|---|---|---|---|
| yolo11x_best | 0.9512 | 0.5209 | 0.9973 | 0.9200 |
| yolo26x_best | 0.9457 | 0.5291 | 0.9683 | 0.9200 |

## Limitations

- Trained on factory environments; performance may degrade in outdoor or non-industrial scenes.
- Heavy occlusion, unusual poses, or extreme camera angles may reduce recall.
- Not evaluated for fairness across demographics; detections are class-level only (no identity).

## Ethical Considerations

These models detect people in workplaces. Deployments should comply with local privacy and labor regulations (e.g., KVKK/GDPR), inform affected workers, and avoid use for individual surveillance or performance tracking.

## License

Released under **AGPL-3.0**, consistent with the [Ultralytics license](https://www.ultralytics.com/legal/agpl-3-0-software-license) of the base models. For commercial licensing of Ultralytics-derived models, see [Ultralytics Licensing](https://www.ultralytics.com/license).

## Citation

```bibtex
@misc{balkontech-worker-detection,
  author = {Kocaaslan, Etem},
  title = {BalkonTech Models: Fine-tuned YOLO for Factory Worker Detection},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/etemkocaaslan/balkontech-models}}
}
```