PPE Sentinel YOLO26

PPE Sentinel is an open-source YOLO26 workplace-safety detector with a browser UI, a local Electron desktop app, a self-hosted HTTP service, and a maintained YOLO dataset import interface.

The bundled incremental model has 28 classes, including Mask, Gloves, Goggles, Safety Vest, Coverall, and Ear Protection. Ear Protection currently has limited representative validation data and must not be used as the sole basis for compliance decisions.

Detailed minimum, recommended, deployment, and training requirements are documented in SYSTEM_REQUIREMENTS.md.

The exact GitHub Actions setup and release procedure is documented in GITHUB_ACTIONS_GUIDE.md.

Clone and setup

Requirements: Node.js 20+, Python 3.10+, and a working C/C++ runtime for the selected PyTorch build. From a clone of this repository:

git clone https://github.com/BulletsHo/ppe-sentinel-yolo26.git
cd ppe-sentinel-yolo26
.\scripts\setup.ps1
npm start

Linux/macOS:

./scripts/setup.sh
npm start

Open http://127.0.0.1:4175/. The local server uses the bundled model at outputs/yolo26-train/ppe-yolo26n-incremental-final/weights/best.pt.

Desktop application

The Electron shell starts an isolated local server, selects an available port, and restricts camera permissions to the application origin.

npm install
npm run desktop

Build installers after installing the build dependencies:

npm run dist:desktop

Artifacts are written to release/. PyInstaller bundles the inference and dataset-import backends, so end users do not need Python installed.

Public HTTPS deployment

Do not expose the development server without authentication. Copy .env.example to .env, set a long random PPE_PASSWORD, and point PPE_DOMAIN at a DNS record for the host. Docker Compose runs PPE Sentinel behind Caddy, which obtains and renews HTTPS certificates:

cp .env.example .env
# edit .env: PPE_DOMAIN, PPE_USERNAME, PPE_PASSWORD
docker compose up -d --build

Remote browser camera access requires HTTPS. The direct Node server also supports PPE_TLS_CERT and PPE_TLS_KEY for deployments that terminate TLS in the application process. PPE_PUBLIC=1 requires both PPE_USERNAME and PPE_PASSWORD unless an explicitly configured authenticated gateway is used.

Useful environment variables:

Variable	Purpose
PPE_HOST / PORT	Bind address and web port
PPE_PUBLIC	Public binding mode; requires Basic Auth
PPE_USERNAME / PPE_PASSWORD	HTTP Basic Auth credentials
PPE_MODEL	Model checkpoint path
PPE_DEVICE / PPE_IMGSZ	cpu, CUDA device, and inference size
PPE_LOG_DIR	Default JSONL detection-log directory
PPE_DATASETS_DIR	Dataset registry directory
PPE_INFERENCE_EXECUTABLE	Packaged inference backend override
PPE_DATASET_IMPORT_EXECUTABLE	Packaged dataset importer override

Dataset maintenance

The UI's maintenance panel accepts a YOLO-format ZIP archive, validates labels, blocks ZIP path traversal, creates a portable data.local.yaml, and records a dataset-manifest.json. Imports never start training automatically.

The same interface is available for automation:

# list imported datasets
curl http://127.0.0.1:4175/api/maintenance/datasets

# import a ZIP archive
curl -X POST -H "Content-Type: application/zip" \
  --data-binary @path/to/dataset.zip \
  "http://127.0.0.1:4175/api/maintenance/datasets/import?name=site-safety"

The CLI remains available for maintenance and CI:

npm run dataset:import -- --source path/to/dataset.zip --name site-safety
npm run dataset:audit -- --data path/to/data.yaml

Training and evaluation

Download a compatible YOLO26 base checkpoint and datasets separately. Dataset archives and working data are intentionally excluded from the repository and release artifacts. After preparing a merged dataset:

npm run evaluate -- --model outputs/yolo26-train/ppe-yolo26n-incremental-final/weights/best.pt \
  --data work/ppe-incremental/data.incremental.yaml --split test
npm run privacy:audit

The repository records the current incremental evaluation in YOLO26_PPE.md. Treat those metrics as dataset-specific benchmarks, not a general safety certification.

Privacy and release checks

npm run privacy:audit scans every file eligible for desktop/Docker release plus the bundled checkpoint for local user paths, email addresses, private-key markers, and common access-token formats. npm run build, npm run pack:desktop, npm run dist:desktop, and the Dockerfile all run this check before packaging.

Training outputs, raw images, ZIP archives, local logs, .env, and nested dataset Git metadata are excluded by .gitignore, .dockerignore, and the Electron file allowlist. Run the audit after replacing the model or adding release files.

License

PPE Sentinel is distributed under the GNU Affero General Public License v3.0 or later. Ultralytics is also AGPL-licensed; see THIRD_PARTY_NOTICES.md and LICENSE before redistributing modified network deployments.

The DOI of this code repository

10.13140/RG.2.2.31556.80008

Sources of training datasets for this system

Dataset 1: Construction Site Safety Image Dataset (Roboflow)

https://www.kaggle.com/datasets/snehilsanyal/construction-site-safety-image-dataset-roboflow/data

Dataset 2: Construction Site Safety Computer Vision Model

https://universe.roboflow.com/roboflow-universe-projects/construction-site-safety/dataset/30

Dataset 3: Safety Goggles - PPE Computer Vision Dataset

https://universe.roboflow.com/database-sjrvw/safety-goggles---ppe/dataset/1

Dataset 4: PPE Dataset for Workplace Computer Vision Dataset

https://universe.roboflow.com/siabar/ppe-dataset-for-workplace/dataset/1

Dataset 5: Coverall Computer Vision Dataset

https://universe.roboflow.com/khroos/coverall-cvebt/dataset/2