README.md

---
license: agpl-3.0
library_name: ultralytics
language:
  - en
tags:
  - object-detection
  - yolo26
  - yolo11
  - ultralytics
  - pytorch
  - onnx
  - construction-safety
  - hazard-detection
datasets:
  - custom
---

# Construction-Hazard-Detection

YOLO-based (primarily YOLO26) models for construction-site hazard detection. These models detect:

- Workers without helmets and/or safety vests
- Workers near machinery or vehicles
- Workers in restricted areas (derived from safety cone clustering)
- Machinery/vehicles near utility poles

This repository provides ready-to-use weights in PyTorch (.pt) and ONNX (.onnx) formats, a demo image, and the class label mapping for easy integration.

👉 For the full end-to-end system (APIs, web UI, training, evaluation, data tools), see the main project: https://github.com/yihong1120/Construction-Hazard-Detection

![demo](./data/examples/demo.jpg)

## Labels

Index-to-name mapping used across all provided models (also in `class_names.txt`):

```
0: Hardhat
1: Mask
2: NO-Hardhat
3: NO-Mask
4: NO-Safety Vest
5: Person
6: Safety Cone
7: Safety Vest
8: Machinery
9: Utility Pole
10: Vehicle
```

## Available models

Models are organized by model family (e.g., YOLO26 vs YOLO11).

### YOLO26 (Recommended)

- PyTorch (Ultralytics):
  - `models/yolo26/pt/yolo26n.pt`
  - `models/yolo26/pt/yolo26s.pt`
  - `models/yolo26/pt/yolo26m.pt`
  - `models/yolo26/pt/yolo26l.pt`
  - `models/yolo26/pt/yolo26x.pt`
- ONNX:
  - `models/yolo26/onnx/yolo26n.onnx`
  - `models/yolo26/onnx/yolo26s.onnx`
  - `models/yolo26/onnx/yolo26m.onnx`
  - `models/yolo26/onnx/yolo26l.onnx`
  - `models/yolo26/onnx/yolo26x.onnx`

### YOLO11 (Legacy)

- PyTorch (Ultralytics):
  - `models/yolo11/pt/yolo11n.pt`
  - `models/yolo11/pt/yolo11s.pt`
  - `models/yolo11/pt/yolo11m.pt`
  - `models/yolo11/pt/yolo11l.pt`
  - `models/yolo11/pt/yolo11x.pt`
- ONNX:
  - `models/yolo11/onnx/yolo11n.onnx`
  - `models/yolo11/onnx/yolo11s.onnx`
  - `models/yolo11/onnx/yolo11m.onnx`
  - `models/yolo11/onnx/yolo11l.onnx`
  - `models/yolo11/onnx/yolo11x.onnx`

Large binaries are tracked with Git LFS.

## Quick start

### A) Ultralytics (PyTorch)

```python
from ultralytics import YOLO

# Load a model (choose the variant that fits your needs)
model = YOLO("models/yolo26/pt/yolo26x.pt")

# Inference on the demo image
results = model("data/examples/demo.jpg", imgsz=640, conf=0.25)

# Parse results (first image)
res = results[0]
boxes = res.boxes  # xyxy, confidence, class
for xyxy, conf, cls_id in zip(boxes.xyxy.tolist(), boxes.conf.tolist(), boxes.cls.tolist()):
    print(xyxy, conf, int(cls_id))
```

CLI option:

```bash
yolo predict model=models/yolo26/pt/yolo26x.pt source=data/examples/demo.jpg imgsz=640 conf=0.25
```

### B) ONNX Runtime

```python
import cv2
import numpy as np
import onnxruntime as ort

# Load and preprocess image to 640x640
img = cv2.imread("data/examples/demo.jpg")
img_rgb = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
size = 640
inp = cv2.resize(img_rgb, (size, size)).astype(np.float32) / 255.0
inp = np.transpose(inp, (2, 0, 1))[None, ...]  # 1x3x640x640

# Run ONNX model
session = ort.InferenceSession("models/yolo26/onnx/yolo26x.onnx", providers=["CPUExecutionProvider"])
input_name = session.get_inputs()[0].name
outputs = session.run(None, {input_name: inp})

pred = outputs[0]  # Typically (1, N, no)
print(pred.shape)
```

Post-processing (NMS, scaling back to original image) follows standard Ultralytics/YOLO routines.

## File structure

```
.
├─ README.md
├─ LICENSE
├─ models/
│  ├─ yolo26/
│  │  ├─ pt/
│  │  │  ├─ yolo26n.pt
│  │  │  ├─ yolo26s.pt
│  │  │  ├─ yolo26m.pt
│  │  │  ├─ yolo26l.pt
│  │  │  └─ yolo26x.pt
│  │  └─ onnx/
│  │     ├─ yolo26n.onnx
│  │     ├─ yolo26s.onnx
│  │     ├─ yolo26m.onnx
│  │     ├─ yolo26l.onnx
│  │     └─ yolo26x.onnx
│  └─ yolo11/
│     ├─ pt/
│     │  ├─ yolo11n.pt
│     │  └─ ...
│     └─ onnx/
│        ├─ yolo11n.onnx
│        └─ ...
├─ data/
│  └─ examples/
│     └─ demo.jpg
└─ class_names.txt
```

## Intended use and limitations

- Intended for research and prototyping in construction safety monitoring.
- Performance depends on camera viewpoint, lighting, occlusion, and domain gap.
- For production, evaluate thoroughly on your target environment and consider rule-based filters and tracking.

## Acknowledgements and sources

- Main project and docs: https://github.com/yihong1120/Construction-Hazard-Detection
- Dataset concept inspired by Roboflow construction safety datasets with extended annotations.
- Roboflow dataset: https://app.roboflow.com/object-detection-qn97p/construction-hazard-detection
- Models trained/exported using Ultralytics YOLO.

## License

This repository is distributed under the AGPL-3.0 license. See `LICENSE` for details and ensure compliance, especially for networked deployments.