╭──────────────────────────────────────────────────────────────╮
│ │
│ AI applied to keeping people alive at work. │
│ │
│ PPE detection · incident prediction · safety copilots │
│ datasets · papers · standards · production tools │
│ │
│ maintained by an HSE engineer who also writes the code │
│ │
╰──────────────────────────────────────────────────────────────╯Most "awesome AI" lists are written by engineers who have never been on a site. This one is not. Everything here is judged on one question: would it actually have stopped an accident?
- PPE detection
- Datasets
- Papers
- Safety copilots and LLM tools
- Standards and regulations
- Where AI still fails
- Contributing
Computer vision that checks whether a worker is wearing a hard hat, vest, gloves or a respirator. This is the most mature use of AI in safety.
| Project | What it gives you | Notes |
|---|---|---|
| ZijianWang-ZW/PPE_detection | Real-time YOLO PPE detection + open dataset | YOLOv5x reported 86.55% mAP; YOLOv5s ~52 FPS. Good speed/accuracy trade-off study. |
| ahmadmughees/SH17dataset | SH17 dataset + benchmarked weights | YOLOv8/v9/v10 weights provided. YOLOv9-e reported >70.9%. Best starting point for benchmarking. |
| VoxDroid/Construction-Site-Safety-PPE-Detection | End-to-end YOLOv8 + Flask dashboard | Training notebooks, pretrained weights, compliance reporting. Closest thing to a deployable system. |
| Vinayakmane47/PPE_detection_YOLO | Simple YOLO PPE detector | Small and readable. Good if you are learning. |
| BrunoCestari/PPE-Detection | YOLOv8 vs Faster R-CNN comparison | Useful if you have to justify a model choice to a client. |
| Ultralytics Construction-PPE docs | Official dataset config + training guide | Fastest path from zero to a trained model. |
Typical classes: Hardhat Mask NO-Hardhat NO-Mask NO-Safety Vest
Person Safety Cone Safety Vest machinery vehicle
| Dataset | Size | Domain |
|---|---|---|
| SH17 | Large, 17 classes | General industrial PPE |
| Construction Site Safety (Roboflow) | ~2,800 annotated images | Construction sites |
| Ultralytics Construction-PPE | Ready YOLO format | Construction PPE |
GitHub topic: ppe-detection |
Many | Browse for niche cases |
GitHub topic: construction-safety |
Many | Wider site-safety work |
Warning on datasets. Nearly all public PPE datasets are collected in daylight, in dry weather, with workers facing the camera. Your night shift, your rain, and your worker bent inside a confined space are not in there. Fine-tune on your own footage before you trust any number.
| Paper | Why it matters |
|---|---|
| PPE detector: a YOLO-based architecture for construction sites | Peer-reviewed baseline for YOLO-based PPE detection. Cite this one. |
More papers being added — PRs very welcome, especially non-English research.
LLMs are good at the paperwork side of safety: risk assessments, method statements, toolbox talks, incident write-ups, and answering "what does the regulation actually say".
| Tool | Use |
|---|---|
| Rifa AI | RAG chatbot pattern (Vercel + Groq) you can point at your own safety manuals. |
| Business AI Chatbot | Clone-per-site RAG bot — load your SOPs, staff ask it questions in plain language. |
| Local LLMs via Ollama | Run offline on site with no internet and no data leaving the company. |
Good LLM safety jobs: drafting risk assessments · summarising incident reports · translating safety instructions for migrant workers · answering "which PPE for this chemical" from an SDS · turning a photo of a hazard into a written observation.
Bad LLM safety jobs: deciding whether a confined space is safe to enter · signing off a permit to work · calculating an exposure limit from memory · anything where a hallucinated number gets someone killed.
| Source | Use |
|---|---|
| HSE UK | UK guidance, EH40 workplace exposure limits, RIDDOR reporting. |
| OSHA | US standards and enforcement data. |
| ILO | International labour safety conventions. |
| NEBOSH | The qualification most safety practitioners actually hold. |
| ISO 45001 | Occupational health and safety management systems. |
Read this section before you sell an AI safety system to anybody.
- A detection is not a control. Spotting a missing hard hat after the fact is monitoring, not prevention. In the hierarchy of control it sits at the very bottom, next to PPE itself.
- False negatives are silent. A model that misses 8% of violations looks great on a dashboard and terrible at an inquest.
- Camera angles decide accuracy, not the model. Most real deployments fail on mounting height and glare, not on mAP.
- Workers game it. If the camera only checks the gate, the helmet comes off after the gate.
- Surveillance is a safety hazard too. A workforce that feels watched stops reporting near-misses, and near-miss reporting is worth more than any model.
- Regulators want records, not predictions. Whatever you build must export evidence a human inspector can read.
Open a PR. One entry per PR, and include one line on what it would have prevented. Entries that are only a demo with no dataset, no weights and no paper will be declined.
CC0 1.0 — public domain. Use it however you like.