Real-time face recognition across live CCTV streams using a distributed, 3-service architecture. Built to scale CV-heavy workloads without blocking the API layer.
┌─────────────────────────────────────────────────────────────────┐
│ CLIENT LAYER │
│ React Dashboard ←→ Socket.IO (real-time alerts) │
└───────────────────────────┬─────────────────────────────────────┘
│ REST API / WebSocket
┌───────────────────────────▼─────────────────────────────────────┐
│ SERVICE 1: Django API Server │
│ • JWT Auth (access + refresh rotation) │
│ • 15+ REST endpoints across 6 Django apps │
│ • MJPEG camera stream ingestion │
│ • Role-based access: Citizen / Police / Admin │
│ • Case lifecycle management │
└──────────┬────────────────────────────────────────┬─────────────┘
│ Enqueue frames via Redis │
┌──────────▼─────────────┐ ┌────────────▼────────────┐
│ SERVICE 2: Celery │ │ SERVICE 3: Node.js │
│ Worker Pool (async) │ │ Socket.IO Server │
│ │ │ │
│ • Batch frames (20s) │ │ • Real-time alert │
│ • dlib 128-d face │ │ push to officers │
│ embedding extraction │ │ • Sub-second delivery │
│ • Configurable match │◄─ alerts ───►│ • Room-based channels │
│ distance threshold │ │ per police station │
│ • Emit match events │ │ │
└─────────────────────────┘ └─────────────────────────┘
│
┌──────────▼─────────────┐
│ PostgreSQL │
│ • Missing persons DB │
│ • Face embedding store │
│ • Case + alert logs │
└─────────────────────────┘
Why 3 services? Face recognition via dlib is CPU-intensive. Running it synchronously would block the Django request cycle under concurrent CCTV loads. The Celery worker pool decouples compute from the API, while the Node.js Socket.IO service handles real-time push without long-polling overhead on Django.
| Feature | Details |
|---|---|
| Real-time face matching | dlib 128-dimensional embeddings with configurable distance thresholds |
| Async pipeline | Frames batched every 20s → Redis queue → Celery workers (no API blocking) |
| Live CCTV streaming | Full MJPEG camera stream ingestion and processing |
| Sub-second alerts | Socket.IO push to police officers the moment a match is confirmed |
| Role-based multi-tenancy | Citizen / Police / Admin with JWT refresh-token rotation |
| Case management | Full case lifecycle: register → investigate → resolve |
| Alert verification | Officers confirm/reject matches before escalation |
| Layer | Technology |
|---|---|
| API Server | Django 4, Django REST Framework |
| CV Engine | OpenCV 4, dlib (128-d face embeddings) |
| Async Workers | Celery 5 + Redis (task queue & frame store) |
| Real-time | Node.js + Socket.IO |
| Frontend | React 18 |
| Database | PostgreSQL 15 |
| Auth | JWT with refresh-token rotation |
- Python 3.10+
- Node.js 18+
- PostgreSQL 15+
- Redis 7+
cmake(required for dlib compilation)
git clone https://github.com/bajpaisatvic/Deep_Vision.git
cd Deep_Vision# Create virtual environment
python -m venv venv
source venv/bin/activate # Windows: venv\Scripts\activate
# Install dependencies
pip install -r requirements.txt
# Configure environment variables
cp .env.example .env
# Edit .env: DATABASE_URL, REDIS_URL, SECRET_KEY, JWT_SECRET
# Run migrations
python manage.py migrate
# Create superuser
python manage.py createsuperuser
# Start Django server
python manage.py runserver# In a new terminal (venv activated)
celery -A config worker --loglevel=info --concurrency=4cd deepvision-realtime
npm install
node index.jscd deepvision-dashboard
npm install
npm run devDeep_Vision/
├── accounts/ # User auth: JWT, role management (Citizen/Police/Admin)
├── alerts/ # Alert creation, verification, and escalation logic
├── cameras/ # CCTV feed ingestion, MJPEG stream handling
├── cases/ # Missing person case lifecycle management
├── config/ # Django settings, Celery config, URL routing
├── deepvision-dashboard/ # React frontend
├── deepvision-realtime/ # Node.js + Socket.IO real-time alert server
└── requirements.txt
1. Camera feed → Django ingests MJPEG frames
2. Frames queued to Redis every 20 seconds (batching reduces redundant processing)
3. Celery worker dequeues batch → runs dlib face detection → extracts 128-d embeddings
4. Embedding compared against missing persons DB using configurable distance threshold
5. On match → alert event emitted to Node.js Socket.IO server
6. Socket.IO pushes alert to the assigned police station room in < 1 second
7. Officer receives alert with face match confidence, camera location, timestamp
8. Officer verifies or rejects the match → case status updated in PostgreSQL
- JWT access + refresh token rotation (tokens invalidated on refresh)
- Role-based route guards — Citizens cannot access police/admin endpoints
- Face embeddings stored as float arrays, never as raw images in the alert log
- Camera credentials stored in environment variables, never hardcoded
| Module | Endpoints | Description |
|---|---|---|
| Auth | POST /api/auth/register, /login, /refresh |
JWT-based auth |
| Cases | GET/POST /api/cases/, PATCH /api/cases/:id/ |
Missing person case CRUD |
| Cameras | GET /api/cameras/, POST /api/cameras/stream/ |
Camera registration & streaming |
| Alerts | GET /api/alerts/, PATCH /api/alerts/:id/verify/ |
Alert retrieval & verification |
| Admin | /api/admin/users/, /api/admin/stats/ |
System stats, user management |