Autonomous helmet enforcement — ESP32-CAM runs Edge Impulse AI at the edge, flags violations, extracts number plates via OCR, and logs every offender to a live web dashboard. No manual intervention. Just power and Wi-Fi.
Live Dashboard: helmut-mhzc.onrender.com
┌─────────────────────────────────────────────────────────────────────┐
│ ESP32-CAM (Edge) │
│ │
│ Boot → WiFiManager portal → Connect Wi-Fi │
│ │ │
│ Every 2 seconds: │
│ 1. Capture QVGA RGB888 frame │
│ 2. Run Edge Impulse model → helmet/no-helmet score │
│ 3. Score ≥ 0.75 → VIOLATION │
│ 4. Recapture VGA JPEG → POST to server │
│ 5. Display result + plate on SSD1306 OLED │
└──────────────────────────┬──────────────────────────────────────────┘
│ HTTP POST /violation
│ multipart: plate_crop (JPEG) + confidence
▼
┌─────────────────────────────────────────────────────────────────────┐
│ Flask Server (Render.com) │
│ │
│ 1. Receive JPEG + confidence │
│ 2. Save snapshot to static/snapshots/ │
│ 3. POST to OCR.space API → extract plate text │
│ 4. Regex match Indian plate format (e.g. MH12AB1234) │
│ 5. INSERT into SQLite offenders table │
│ 6. Push SSE event to all open dashboard browsers │
│ 7. Return {"plate": "MH12AB1234"} to ESP32 │
└──────────────────────────┬──────────────────────────────────────────┘
│ SSE push
▼
┌─────────────────────────────────────────────────────────────────────┐
│ Live Dashboard (Any Browser) │
│ │
│ New row appears instantly — no page refresh │
│ Plate number │ Timestamp │ Confidence │ Snapshot thumbnail │
│ Total count + today count + uptime │
└─────────────────────────────────────────────────────────────────────┘
| Component | Model / Spec | Role | Cost |
|---|---|---|---|
| Microcontroller + Camera | ESP32-CAM (AI-Thinker, OV2640) | Edge AI inference + image capture | ₹450 |
| OLED Display | 1.3" SSD1306, I2C | Live result display | ₹200 |
| FTDI Programmer | CH340 / CP2102 | Flash ESP32-CAM | ₹250 |
| Breadboard + Jumpers | Half-size + DuPont cables | Prototyping | ₹150 |
| Power Supply | 5V / 2A USB adapter | Stable power for CAM | ₹280 |
| Total | ₹1,330 |
OLED SCL → ESP32-CAM IO15
OLED SDA → ESP32-CAM IO14
OLED VCC → 3.3V
OLED GND → GND
helmut/
├── esp32/
│ └── helmet_violation/
│ ├── helmet_violation.ino # Main Arduino sketch
│ └── sxmnath-project-1_inferencing.h # Edge Impulse model header
│
├── server/
│ ├── app.py # Flask backend — receives violations, serves dashboard
│ ├── ocr.py # OCR.space API wrapper + plate regex parser
│ ├── database.py # SQLite init, insert, query
│ ├── requirements.txt
│ ├── .python-version # Pins Python 3.11.9 for Render
│ └── templates/
│ └── dashboard.html # Live SSE dashboard
│
├── simulator/
│ └── simulate_esp32.py # Send test images to server without hardware
│
├── test_images/ # Drop motorbike/plate images here for testing
├── render.yaml # Render deployment config
├── .gitignore
└── README.md
First boot — creates a Wi-Fi hotspot called HelmetMonitor-Setup. Connect your phone to it, enter your Wi-Fi credentials in the captive portal. Credentials are saved to flash — every reboot after connects automatically without any intervention.
Detection loop — every 2 seconds, captures a QVGA frame in RGB888 format and feeds it to the Edge Impulse model via run_classifier(). The model returns confidence scores for with_helmet and without_helmet.
Violation — if without_helmet score ≥ 0.75, the camera switches to JPEG mode, captures a VGA frame, and HTTP POSTs it to the Flask server as a multipart form upload along with the confidence score.
OLED — shows HELMET OK with confidence, or VIOLATION! Plate: XXXXXX when a violation is confirmed. The plate number is returned by the server after OCR.
Receives the JPEG and saves it as a timestamped snapshot. Sends the image to OCR.space (free, 25k calls/month, no card needed). Parses the returned text with a regex for Indian plate format (MH12AB1234). Writes the record to SQLite. Pushes a Server-Sent Event to every open browser — the dashboard row appears instantly without any page refresh.
cd server
pip install -r requirements.txtCreate server/.env:
OCR_SPACE_API_KEY=K8xxxxxxxxxxxxxxx
Get a free key (no card) at ocr.space/ocrapi.
# Mac/Linux
export OCR_SPACE_API_KEY=K8xxxxxxxxxxxxxxx
python app.py
# Windows PowerShell
$env:OCR_SPACE_API_KEY="K8xxxxxxxxxxxxxxx"
python app.pyOpen http://localhost:5000 — dark dashboard, zero violations.
cd simulator
python simulate_esp32.py --image ../test_images/no_helmet_with_plate.jpgLoop mode — cycles through all images in test_images/ every 5 seconds:
python simulate_esp32.py --loop --delay 5
python simulate_esp32.py --loop --random --delay 3Expected output:
[OK] no_helmet_with_plate.jpg plate=MH12AB1234 ts=20260328_154322_001
Dashboard at http://localhost:5000 updates live — new row appears without any page refresh.
Open esp32/helmet_violation/helmet_violation.ino in Arduino IDE. Install the Edge Impulse library by adding the .zip exported from your Edge Impulse project. Install WiFiManager by tzapu via Library Manager.
Board settings:
| Setting | Value |
|---|---|
| Board | AI Thinker ESP32-CAM |
| Partition Scheme | Huge APP (3MB No OTA) |
| CPU Frequency | 240MHz |
| Upload Speed | 115200 |
Update SERVER_URL in the sketch to your Render URL, upload, open Serial Monitor at 115200 baud.
On first boot the ESP32 creates a hotspot called HelmetMonitor-Setup — connect your phone to it, enter Wi-Fi credentials, and it connects and starts detecting automatically.
render.yaml at the repo root handles the full deployment config. Steps:
- Go to render.com → New → Web Service → connect this repo
- Render auto-detects
render.yaml— confirm Root Directory isserver - Add one environment variable:
OCR_SPACE_API_KEY = K8xxxxxxxxxxxxxxx - Click Deploy — build completes in ~2 minutes
No heavy ML dependencies on the server. OCR is a pure HTTP call to OCR.space, so the entire server stack is under 50MB.
| Layer | Technology |
|---|---|
| Firmware | C++ (Arduino framework, ESP32-CAM) |
| Edge AI model | Edge Impulse (TFLite Micro, exported as .h header) |
| Wi-Fi provisioning | WiFiManager by tzapu (captive portal, no hardcoded credentials) |
| Backend | Python, Flask |
| OCR | OCR.space API (free tier, 25k calls/month) |
| Database | SQLite |
| Live updates | Server-Sent Events (SSE) — zero polling, instant push |
| Frontend | HTML, CSS, Vanilla JS |
| Hosting | Render.com (free tier) |
Ephemeral storage on Render free tier — snapshots and the SQLite database reset on every redeploy. For persistent storage, swap SQLite for Render PostgreSQL (free tier available) and snapshots for Cloudinary.
Render cold start — free tier sleeps after 15 minutes of inactivity. The ESP32 sends a keepalive GET to /health every 10 minutes while powered on to prevent this.
OCR accuracy — depends on image quality and plate visibility. Works best when the camera is mounted at a fixed choke point (gate, speed bump) 2–3 metres high, angled so both rider and plate are in frame simultaneously.
Two-frame capture — helmet inference runs on an RGB frame; plate POST uses a separately captured JPEG frame. If the rider moves between the two captures (~200ms apart), the plate crop may be blurry or empty. A single-frame dual-crop approach would eliminate this — a future improvement.
MIT