| Scanner | Stoffdetails · Substance detail | Absperrradius · Isolation distance |
|---|---|---|
 |
 |
 |
Die Warntafel im ersten Bild ist synthetisch erzeugt, mit dem Generator aus scripts/.
The placard in the first image is synthetic, produced by the generator in scripts/.
Offline-Gefahrstoff-Assistent für Feuerwehren. Kamera auf eine orange ADR-Warntafel halten: Die App liest Kemler-Zahl und UN-Nummer, identifiziert den Stoff und zeigt die Sofortmaßnahmen. Alles auf dem Gerät, ohne Netzverbindung.
Der Detektor ist ein eigens dafür trainiertes YOLO26-nano-Modell, nach TFLite exportiert und auf dem Gerät ausgeführt. Die Stoffdaten stammen aus ADR Tabelle A, die Einsatzmaßnahmen von den CEFIC ERICards.
Plattform: Android. iOS wird nicht unterstützt, siehe Lizenz.
Die App ist auf Deutsch, sie ist für Feuerwehren in Deutschland, Österreich und der Schweiz gebaut. Code und Dokumentation sind auf Englisch.
Diese App ist ein Hilfsmittel. Sie ersetzt weder Ausbildung noch Erfahrung noch die Beurteilung der Einsatzkräfte vor Ort; es entscheidet die Einsatzleitung. Alle Daten sind ohne Gewähr und können veraltet oder falsch sein. Alles, worauf es ankommt, ist gegen die geltenden Vorschriften zu prüfen; im Zweifel ist ein Fachberater Gefahrgut hinzuzuziehen.
Das ist ein Hobbyprojekt, kein zertifiziertes Produkt.
- Scanner: Live-Erkennung oranger ADR-Warntafeln, OCR von Kemler-Zahl und UN-Nummer auf dem Gerät, Stoffidentifikation. Trägt eine UN-Nummer mehrere Einträge, löst die Kemler-Zahl auf, welcher davon auf dem Fahrzeug steht.
- Suche nach UN-Nummer, Stoffname oder Kemler-Zahl
- ERICards mit Sofortmaßnahmen, Gefahren, Schutzausrüstung, Löschmitteln
- Absperrradien nach ERG 2024, mit Karte
- Rechner und Merkhilfen: Atemluft, ppm-Umrechnung, Volumenabschätzung, Beaufort-Skala, GAMS-Checkliste
Die technische Dokumentation ab Getting started ist auf Englisch.
An offline hazmat assistant for firefighters. Point the camera at an orange ADR placard and the app reads the Kemler and UN number, identifies the substance and shows the immediate measures, all on the device and without a network connection.
The detector is a YOLO26-nano model trained specifically for this, exported to TFLite and running on-device. The substance data comes from ADR Table A, the emergency measures from CEFIC ERICards.
Platform: Android. iOS is not supported, see Licence.
The app language is German, since it is built for fire services in Germany, Austria and Switzerland. Code and documentation are English.
This app is an aid. It does not replace training, experience or the judgement of the people on scene, and the incident commander decides. All data is provided without warranty and may be outdated or wrong. Verify anything that matters against the current regulations and involve a hazmat adviser.
This is a hobby project, not a certified product.
- Scanner: live detection of orange ADR placards, on-device OCR of the Kemler and UN number, substance identification. Where a UN number has several entries, the Kemler number resolves which one is on the truck.
- Search by UN number, substance name or Kemler number
- ERICards with immediate measures, hazards, protective equipment, firefighting media
- Isolation distances to ERG 2024, with a map
- Calculators and reference cards: breathing air, ppm conversion, volume estimation, Beaufort scale, GAMS checklist
- Node.js 20 or newer, npm
- Python 3.12 or newer
- Android device with API 28 or newer, developer mode enabled
- JDK 17 and the Android SDK for local builds
The substance data ships with the repository. The ERICards do not: "free for fire services" is a permission to use, not a permission for us to redistribute, so you fetch your own copy.
python3 scripts/parse-ericards.py
Around 2,300 cards in roughly 40 minutes. It runs unattended and can be resumed, so an interruption costs nothing. Until it has run the app works, but a substance shows no emergency measures.
Please leave the rate limit alone. The scraper waits half a second between requests because the site belongs to someone else and carries no advertising to pay for the traffic.
Which terms apply to both datasets, and what the attribution has to carry, is in docs/data-sources.md.
cd apps/mobile
npm install
./build-android.sh release --device
Always build onto a real device. The camera and the GPU frame processing do not work in an
emulator, and with one configured expo run:android picks it without asking. A failed
emulator start still exits zero, so that failure reads as a successful build.
./build-android.sh without an argument builds the debug variant, which is the faster loop
while developing. Use the release variant to check anything that behaves differently in a
shipped app, the model loading above all.
cd apps/mobile && npm test # logic and components
python3 -m unittest discover -s scripts -p 'test_*.py'
The database tests run against a real in-memory SQLite instance including the full text index, not against mocks.
scripts/ holds the whole chain: synthetic training data, YOLO26-nano training and the
TFLite export. The trained model is in the repository, so this is optional. See
docs/training.md.
Substance data: ADR Table A from the Bundesanstalt für Materialforschung und -prüfung (BAM), Datenbank GEFAHRGUT, under Datenlizenz Deutschland - Namensnennung - Version 2.0. The attribution the licence requires is carried in the app and spelled out in docs/data-sources.md.
ERICards: CEFIC, free for fire services, fetched by you rather than redistributed here.
Isolation distances: ERG 2024 (PHMSA / Transport Canada), public domain.
Breathing air: FwDV 7.
| docs/architecture.md | how the app is put together and why offline |
| docs/scanner-pipeline.md | camera frame to substance, including what failed |
| docs/data-sources.md | where the data comes from and under which terms |
| docs/training.md | retraining the detector |
AGPL-3.0. See LICENSE.
The detector is built with Ultralytics, which is AGPL-3.0 itself and treats trained weights as derived works. The whole project follows suit, which also settles the question rather than leaving it open.
One consequence worth knowing: the GPL family conflicts with Apple's App Store terms, so this cannot be distributed on iOS. Android and self-built installs are unaffected.