feat: initial BantayCam scaffold

- FastAPI backend with async SQLAlchemy - Camera RTSP management (add, start, stop) - Vehicle detection (YOLO + fast-alpr) - Type: car, motorcycle, truck, jeepney - Color detection (HSV) - License plate OCR - Motorcycle person count - Face detection + InsightFace ArcFace embedding - pgvector identity grouping (auto-cluster same face) - Vehicle + person movement trail APIs - Docker Compose with pgvector/pg16 - Models: Camera, VehicleIdentity, VehicleEvent, PersonIdentity, FaceEvent
2026-03-12 12:11:02 +08:00
commit 7c43e4580d
24 changed files with 1650 additions and 0 deletions
--- a/.env.example
+++ b/.env.example
@@ -0,0 +1,23 @@
+# BantayCam Environment Variables
+# Copy to .env and fill in your values
+
+# Database
+DATABASE_URL=postgresql+asyncpg://bantaycam:bantaycam@localhost:5432/bantaycam
+
+# Storage
+STORAGE_TYPE=local
+STORAGE_PATH=./snapshots
+
+# AI Models
+YOLO_MODEL=yolov8n.pt
+FACE_MODEL=buffalo_l
+GPU_DEVICE=0              # Set to -1 for CPU-only
+
+# Processing tuning
+PROCESS_EVERY_N_FRAMES=5
+FACE_SIMILARITY_THRESHOLD=0.5
+PLATE_CONFIDENCE_THRESHOLD=0.6
+
+# Alerts (optional)
+TELEGRAM_BOT_TOKEN=
+TELEGRAM_CHAT_ID=
--- a/README.md
+++ b/README.md
@@ -0,0 +1,107 @@
+# 🎥 BantayCam
+
+**Integrated License Plate + Face Recognition for Philippine CCTV Infrastructure**
+
+Built with Python, FastAPI, YOLOv8, InsightFace, and PostgreSQL.
+
+---
+
+## Features
+
+- 📡 **RTSP Camera Management** — Add any IP camera by RTSP URL, name it, start/stop anytime
+- 🚗 **Vehicle Detection** — Detects car, motorcycle, jeepney, truck; identifies type + color
+- 🔢 **License Plate Recognition** — Reads PH plates; logs every appearance with timestamp + camera
+- 🏍️ **Motorcycle Person Count** — Counts riders on motorcycles
+- 👤 **Face Detection** — Detects and embeds every face using InsightFace ArcFace
+- 🔗 **Identity Grouping** — Auto-groups same face across events using pgvector similarity search
+- 🗺️ **Movement Trail** — Query where a face or plate has been seen, on which camera, and when
+- 🏷️ **Labeling** — Register identities with names, mark vehicles as resident/blacklisted
+
+---
+
+## Quick Start
+
+### Requirements
+- Docker + Docker Compose
+- (Optional) NVIDIA GPU + CUDA for real-time processing
+
+### Run with Docker
+
+```bash
+git clone http://192.168.1.159:3000/kibin/bantaycam.git
+cd bantaycam
+cp .env.example .env
+docker-compose up -d
+```
+
+API available at: http://localhost:8000
+Swagger docs: http://localhost:8000/docs
+
+---
+
+## API Endpoints
+
+### Cameras
+| Method | Endpoint | Description |
+|---|---|---|
+| GET | /api/v1/cameras/ | List all cameras |
+| POST | /api/v1/cameras/ | Register new RTSP camera |
+| POST | /api/v1/cameras/{id}/start | Start stream |
+| POST | /api/v1/cameras/{id}/stop | Stop stream |
+| GET | /api/v1/cameras/streams/status | Live stream status |
+
+### Vehicles
+| Method | Endpoint | Description |
+|---|---|---|
+| GET | /api/v1/vehicles/identities | List all known vehicles |
+| GET | /api/v1/vehicles/identities/{id}/trail | Where was this vehicle seen? |
+| GET | /api/v1/vehicles/events | Recent vehicle detection events |
+| PATCH | /api/v1/vehicles/identities/{id} | Label/whitelist/blacklist a vehicle |
+
+### Persons
+| Method | Endpoint | Description |
+|---|---|---|
+| GET | /api/v1/persons/identities | List all detected identities |
+| GET | /api/v1/persons/identities/{id}/trail | Where was this person seen? |
+| PATCH | /api/v1/persons/identities/{id} | Register name/watchlist |
+| GET | /api/v1/persons/events | Recent face detection events |
+
+---
+
+## Architecture
+
+```
+Camera (RTSP) → StreamWorker → Frame Queue → DetectionPipeline
+                                                    │
+                               ┌────────────────────┤
+                               │                    │
+                        VehicleDetector       FaceService
+                        (YOLO + ALPR)         (InsightFace)
+                               │                    │
+                               └────────────────────┤
+                                                    │
+                                            PostgreSQL + pgvector
+                                            (Events, Identities, Embeddings)
+```
+
+---
+
+## Tech Stack
+
+| Component | Technology |
+|---|---|
+| Backend API | FastAPI + Python 3.11 |
+| Vehicle Detection | YOLOv8 (Ultralytics) |
+| License Plate OCR | fast-alpr (YOLOv9 + ONNX) |
+| Face Recognition | InsightFace ArcFace (buffalo_l) |
+| Database | PostgreSQL 16 + pgvector |
+| Containerization | Docker Compose |
+| Stream Ingestion | OpenCV RTSP |
+
+---
+
+## Development Roadmap
+
+- [ ] Phase 1: Core — RTSP + vehicle logging + face detection ← **YOU ARE HERE**
+- [ ] Phase 2: Multi-stream + SMS/push alerts + gate relay
+- [ ] Phase 3: Web dashboard + mobile app + LTO API integration
--- a/backend/Dockerfile
+++ b/backend/Dockerfile
@@ -0,0 +1,16 @@
+FROM python:3.11-slim
+
+WORKDIR /app
+
+# System deps for OpenCV + InsightFace
+RUN apt-get update && apt-get install -y \
+    libgl1 libglib2.0-0 libgomp1 \
+    && rm -rf /var/lib/apt/lists/*
+
+COPY requirements.txt .
+RUN pip install --no-cache-dir -r requirements.txt
+
+COPY . .
+
+EXPOSE 8000
+CMD ["uvicorn", "app.main:app", "--host", "0.0.0.0", "--port", "8000"]
--- a/backend/app/init.py
+++ b/backend/app/init.py
--- a/backend/app/api/init.py
+++ b/backend/app/api/init.py
--- a/backend/app/api/cameras.py
+++ b/backend/app/api/cameras.py
@@ -0,0 +1,118 @@
+"""
+Camera API — Register, manage, and monitor RTSP cameras.
+"""
+from fastapi import APIRouter, Depends, HTTPException, status
+from sqlalchemy.ext.asyncio import AsyncSession
+from sqlalchemy import select
+from pydantic import BaseModel
+from typing import Optional
+from uuid import UUID
+
+from app.core.database import get_db
+from app.models.camera import Camera, CameraStatus
+from app.services.stream_manager import StreamManager
+from app.workers.pipeline import DetectionPipeline
+
+router = APIRouter(prefix="/cameras", tags=["cameras"])
+
+
+class CameraCreate(BaseModel):
+    name: str
+    rtsp_url: str
+    location: Optional[str] = None
+    description: Optional[str] = None
+    auto_start: bool = True
+
+
+class CameraResponse(BaseModel):
+    id: UUID
+    name: str
+    rtsp_url: str
+    location: Optional[str]
+    description: Optional[str]
+    status: str
+    is_enabled: bool
+
+    class Config:
+        from_attributes = True
+
+
+@router.get("/", response_model=list[CameraResponse])
+async def list_cameras(db: AsyncSession = Depends(get_db)):
+    result = await db.execute(select(Camera).order_by(Camera.name))
+    return result.scalars().all()
+
+
+@router.post("/", response_model=CameraResponse, status_code=status.HTTP_201_CREATED)
+async def add_camera(payload: CameraCreate, db: AsyncSession = Depends(get_db)):
+    """Register a new RTSP camera and optionally start streaming."""
+    # Check duplicate name
+    existing = await db.execute(select(Camera).where(Camera.name == payload.name))
+    if existing.scalar_one_or_none():
+        raise HTTPException(status_code=400, detail="Camera name already exists")
+
+    camera = Camera(
+        name=payload.name,
+        rtsp_url=payload.rtsp_url,
+        location=payload.location,
+        description=payload.description,
+        status=CameraStatus.inactive,
+    )
+    db.add(camera)
+    await db.commit()
+    await db.refresh(camera)
+
+    if payload.auto_start:
+        await _start_camera(camera)
+
+    return camera
+
+
+@router.post("/{camera_id}/start")
+async def start_camera(camera_id: UUID, db: AsyncSession = Depends(get_db)):
+    camera = await _get_camera(db, camera_id)
+    await _start_camera(camera)
+    camera.status = CameraStatus.active
+    await db.commit()
+    return {"message": f"Camera '{camera.name}' started"}
+
+
+@router.post("/{camera_id}/stop")
+async def stop_camera(camera_id: UUID, db: AsyncSession = Depends(get_db)):
+    camera = await _get_camera(db, camera_id)
+    manager = StreamManager.get()
+    await manager.stop_stream(camera_id)
+    camera.status = CameraStatus.inactive
+    await db.commit()
+    return {"message": f"Camera '{camera.name}' stopped"}
+
+
+@router.delete("/{camera_id}", status_code=status.HTTP_204_NO_CONTENT)
+async def delete_camera(camera_id: UUID, db: AsyncSession = Depends(get_db)):
+    camera = await _get_camera(db, camera_id)
+    await StreamManager.get().stop_stream(camera_id)
+    await db.delete(camera)
+    await db.commit()
+
+
+@router.get("/streams/status")
+async def stream_status():
+    """Get real-time status of all active streams."""
+    return StreamManager.get().get_all_status()
+
+
+# --- Helpers ---
+
+async def _get_camera(db: AsyncSession, camera_id: UUID) -> Camera:
+    result = await db.execute(select(Camera).where(Camera.id == camera_id))
+    camera = result.scalar_one_or_none()
+    if not camera:
+        raise HTTPException(status_code=404, detail="Camera not found")
+    return camera
+
+
+async def _start_camera(camera: Camera):
+    manager = StreamManager.get()
+    worker = await manager.start_stream(camera.id, camera.name, camera.rtsp_url)
+    pipeline = DetectionPipeline(camera.id, camera.name, worker)
+    await pipeline.start()
--- a/backend/app/api/persons.py
+++ b/backend/app/api/persons.py
@@ -0,0 +1,130 @@
+"""
+Person API — Face identities and sighting trails.
+See everywhere a face was detected, which cameras, and when.
+"""
+from fastapi import APIRouter, Depends, Query, UploadFile, File
+from sqlalchemy.ext.asyncio import AsyncSession
+from sqlalchemy import select, desc
+from pydantic import BaseModel
+from typing import Optional
+from uuid import UUID
+from datetime import datetime
+
+from app.core.database import get_db
+from app.models.person import PersonIdentity, FaceEvent
+
+router = APIRouter(prefix="/persons", tags=["persons"])
+
+
+class PersonIdentityResponse(BaseModel):
+    id: UUID
+    name: Optional[str]
+    label: Optional[str]
+    is_registered: int
+    is_watchlisted: int
+    thumbnail_path: Optional[str]
+    first_seen_at: Optional[datetime]
+    last_seen_at: Optional[datetime]
+    total_sightings: int
+
+    class Config:
+        from_attributes = True
+
+
+@router.get("/identities", response_model=list[PersonIdentityResponse])
+async def list_identities(
+    name: Optional[str] = Query(None),
+    limit: int = Query(50, le=200),
+    db: AsyncSession = Depends(get_db),
+):
+    q = select(PersonIdentity).order_by(desc(PersonIdentity.last_seen_at)).limit(limit)
+    if name:
+        q = q.where(PersonIdentity.name.ilike(f"%{name}%"))
+    result = await db.execute(q)
+    return result.scalars().all()
+
+
+@router.get("/identities/{identity_id}/trail")
+async def person_trail(identity_id: UUID, db: AsyncSession = Depends(get_db)):
+    """
+    Full sighting trail — every camera this face appeared on and when.
+    Groups by camera to show movement pattern.
+    """
+    result = await db.execute(
+        select(FaceEvent)
+        .where(FaceEvent.identity_id == identity_id)
+        .order_by(desc(FaceEvent.captured_at))
+        .limit(500)
+    )
+    events = result.scalars().all()
+    return {
+        "identity_id": str(identity_id),
+        "total_sightings": len(events),
+        "trail": [
+            {
+                "camera_id": str(e.camera_id),
+                "captured_at": e.captured_at,
+                "similarity_score": e.similarity_score,
+                "face_snapshot_path": e.face_snapshot_path,
+                "snapshot_path": e.snapshot_path,
+                "vehicle_event_id": str(e.vehicle_event_id) if e.vehicle_event_id else None,
+            }
+            for e in events
+        ]
+    }
+
+
+@router.patch("/identities/{identity_id}")
+async def register_identity(
+    identity_id: UUID,
+    name: Optional[str] = None,
+    label: Optional[str] = None,
+    is_watchlisted: Optional[int] = None,
+    notes: Optional[str] = None,
+    db: AsyncSession = Depends(get_db),
+):
+    """Give a name/label to an auto-detected unknown identity."""
+    result = await db.execute(select(PersonIdentity).where(PersonIdentity.id == identity_id))
+    identity = result.scalar_one_or_none()
+    if not identity:
+        from fastapi import HTTPException
+        raise HTTPException(status_code=404, detail="Person identity not found")
+
+    if name is not None:
+        identity.name = name
+        identity.is_registered = 1
+    if label is not None:
+        identity.label = label
+    if is_watchlisted is not None:
+        identity.is_watchlisted = is_watchlisted
+
+    await db.commit()
+    return {"message": "Identity registered"}
+
+
+@router.get("/events")
+async def list_face_events(
+    camera_id: Optional[UUID] = Query(None),
+    identity_id: Optional[UUID] = Query(None),
+    limit: int = Query(100, le=500),
+    db: AsyncSession = Depends(get_db),
+):
+    q = select(FaceEvent).order_by(desc(FaceEvent.captured_at)).limit(limit)
+    if camera_id:
+        q = q.where(FaceEvent.camera_id == camera_id)
+    if identity_id:
+        q = q.where(FaceEvent.identity_id == identity_id)
+    result = await db.execute(q)
+    events = result.scalars().all()
+    return [
+        {
+            "id": str(e.id),
+            "camera_id": str(e.camera_id),
+            "identity_id": str(e.identity_id) if e.identity_id else None,
+            "detection_confidence": e.detection_confidence,
+            "similarity_score": e.similarity_score,
+            "face_snapshot_path": e.face_snapshot_path,
+            "captured_at": e.captured_at,
+        }
+        for e in events
+    ]
--- a/backend/app/api/vehicles.py
+++ b/backend/app/api/vehicles.py
@@ -0,0 +1,139 @@
+"""
+Vehicle API — Query vehicle events and identities.
+See where a plate has been seen across all cameras and when.
+"""
+from fastapi import APIRouter, Depends, Query
+from sqlalchemy.ext.asyncio import AsyncSession
+from sqlalchemy import select, desc
+from pydantic import BaseModel
+from typing import Optional
+from uuid import UUID
+from datetime import datetime
+
+from app.core.database import get_db
+from app.models.vehicle import VehicleEvent, VehicleIdentity
+
+router = APIRouter(prefix="/vehicles", tags=["vehicles"])
+
+
+class VehicleIdentityResponse(BaseModel):
+    id: UUID
+    plate_number: Optional[str]
+    vehicle_type: str
+    color: Optional[str]
+    label: Optional[str]
+    is_whitelisted: int
+    is_blacklisted: int
+    first_seen_at: Optional[datetime]
+    last_seen_at: Optional[datetime]
+    total_sightings: int
+
+    class Config:
+        from_attributes = True
+
+
+class VehicleEventResponse(BaseModel):
+    id: UUID
+    camera_id: UUID
+    plate_number: Optional[str]
+    vehicle_type: str
+    color: Optional[str]
+    person_count: Optional[int]
+    detection_confidence: Optional[float]
+    snapshot_path: Optional[str]
+    plate_snapshot_path: Optional[str]
+    captured_at: datetime
+
+    class Config:
+        from_attributes = True
+
+
+@router.get("/identities", response_model=list[VehicleIdentityResponse])
+async def list_vehicle_identities(
+    plate: Optional[str] = Query(None),
+    vehicle_type: Optional[str] = Query(None),
+    limit: int = Query(50, le=200),
+    db: AsyncSession = Depends(get_db),
+):
+    q = select(VehicleIdentity).order_by(desc(VehicleIdentity.last_seen_at)).limit(limit)
+    if plate:
+        q = q.where(VehicleIdentity.plate_number.ilike(f"%{plate}%"))
+    if vehicle_type:
+        q = q.where(VehicleIdentity.vehicle_type == vehicle_type)
+    result = await db.execute(q)
+    return result.scalars().all()
+
+
+@router.get("/identities/{identity_id}/trail")
+async def vehicle_trail(identity_id: UUID, db: AsyncSession = Depends(get_db)):
+    """
+    Get full sighting trail for a vehicle — every camera it was seen on and when.
+    """
+    result = await db.execute(
+        select(VehicleEvent)
+        .where(VehicleEvent.identity_id == identity_id)
+        .order_by(desc(VehicleEvent.captured_at))
+        .limit(500)
+    )
+    events = result.scalars().all()
+    return {
+        "identity_id": str(identity_id),
+        "total_sightings": len(events),
+        "trail": [
+            {
+                "camera_id": str(e.camera_id),
+                "captured_at": e.captured_at,
+                "plate_number": e.plate_number,
+                "color": e.color,
+                "vehicle_type": e.vehicle_type,
+                "person_count": e.person_count,
+                "snapshot_path": e.snapshot_path,
+            }
+            for e in events
+        ]
+    }
+
+
+@router.get("/events", response_model=list[VehicleEventResponse])
+async def list_vehicle_events(
+    camera_id: Optional[UUID] = Query(None),
+    plate: Optional[str] = Query(None),
+    limit: int = Query(100, le=500),
+    db: AsyncSession = Depends(get_db),
+):
+    q = select(VehicleEvent).order_by(desc(VehicleEvent.captured_at)).limit(limit)
+    if camera_id:
+        q = q.where(VehicleEvent.camera_id == camera_id)
+    if plate:
+        q = q.where(VehicleEvent.plate_number.ilike(f"%{plate}%"))
+    result = await db.execute(q)
+    return result.scalars().all()
+
+
+@router.patch("/identities/{identity_id}")
+async def update_vehicle_identity(
+    identity_id: UUID,
+    label: Optional[str] = None,
+    is_whitelisted: Optional[int] = None,
+    is_blacklisted: Optional[int] = None,
+    notes: Optional[str] = None,
+    db: AsyncSession = Depends(get_db),
+):
+    """Label a vehicle — mark as resident, delivery, blacklisted, etc."""
+    result = await db.execute(select(VehicleIdentity).where(VehicleIdentity.id == identity_id))
+    identity = result.scalar_one_or_none()
+    if not identity:
+        from fastapi import HTTPException
+        raise HTTPException(status_code=404, detail="Vehicle identity not found")
+
+    if label is not None:
+        identity.label = label
+    if is_whitelisted is not None:
+        identity.is_whitelisted = is_whitelisted
+    if is_blacklisted is not None:
+        identity.is_blacklisted = is_blacklisted
+    if notes is not None:
+        identity.notes = notes
+
+    await db.commit()
+    return {"message": "Updated"}
--- a/backend/app/core/init.py
+++ b/backend/app/core/init.py
--- a/backend/app/core/config.py
+++ b/backend/app/core/config.py
@@ -0,0 +1,42 @@
+from pydantic_settings import BaseSettings
+from typing import Optional
+
+
+class Settings(BaseSettings):
+    APP_NAME: str = "BantayCam"
+    APP_VERSION: str = "0.1.0"
+    DEBUG: bool = False
+
+    # Database
+    DATABASE_URL: str = "postgresql+asyncpg://bantaycam:bantaycam@localhost:5432/bantaycam"
+
+    # Storage (MinIO / local)
+    STORAGE_TYPE: str = "local"  # "local" or "minio"
+    STORAGE_PATH: str = "./snapshots"
+    MINIO_ENDPOINT: str = ""
+    MINIO_ACCESS_KEY: str = ""
+    MINIO_SECRET_KEY: str = ""
+    MINIO_BUCKET: str = "bantaycam"
+
+    # AI Models
+    YOLO_MODEL: str = "yolov8n.pt"              # Vehicle + person detection
+    PLATE_MODEL: str = "yolov8n.pt"             # Plate detection (custom PH)
+    FACE_MODEL: str = "buffalo_l"               # InsightFace model
+    GPU_DEVICE: int = 0                          # -1 for CPU, 0 for first GPU
+
+    # Processing
+    PROCESS_EVERY_N_FRAMES: int = 5             # Skip frames for performance
+    FACE_SIMILARITY_THRESHOLD: float = 0.5      # Cosine distance threshold
+    PLATE_CONFIDENCE_THRESHOLD: float = 0.6
+    DETECTION_CONFIDENCE_THRESHOLD: float = 0.5
+
+    # Alerts
+    TELEGRAM_BOT_TOKEN: Optional[str] = None
+    TELEGRAM_CHAT_ID: Optional[str] = None
+
+    class Config:
+        env_file = ".env"
+        case_sensitive = True
+
+
+settings = Settings()
--- a/backend/app/core/database.py
+++ b/backend/app/core/database.py
@@ -0,0 +1,34 @@
+from sqlalchemy.ext.asyncio import create_async_engine, AsyncSession, async_sessionmaker
+from sqlalchemy.orm import DeclarativeBase
+from app.core.config import settings
+
+
+engine = create_async_engine(
+    settings.DATABASE_URL,
+    echo=settings.DEBUG,
+    pool_pre_ping=True,
+)
+
+AsyncSessionLocal = async_sessionmaker(
+    engine,
+    class_=AsyncSession,
+    expire_on_commit=False,
+)
+
+
+class Base(DeclarativeBase):
+    pass
+
+
+async def get_db():
+    async with AsyncSessionLocal() as session:
+        try:
+            yield session
+        finally:
+            await session.close()
+
+
+async def init_db():
+    """Create all tables on startup."""
+    async with engine.begin() as conn:
+        await conn.run_sync(Base.metadata.create_all)
--- a/backend/app/main.py
+++ b/backend/app/main.py
@@ -0,0 +1,91 @@
+"""
+BantayCam — Integrated License Plate & Face Recognition for Philippine CCTV
+FastAPI Application Entry Point
+"""
+from contextlib import asynccontextmanager
+import logging
+
+from fastapi import FastAPI
+from fastapi.middleware.cors import CORSMiddleware
+from fastapi.staticfiles import StaticFiles
+from pathlib import Path
+
+from app.core.config import settings
+from app.core.database import init_db
+from app.services.vehicle_detector import vehicle_detector
+from app.services.face_service import face_service
+from app.services.stream_manager import StreamManager
+from app.api import cameras, vehicles, persons
+
+logging.basicConfig(
+    level=logging.INFO,
+    format="%(asctime)s [%(levelname)s] %(name)s: %(message)s"
+)
+logger = logging.getLogger(__name__)
+
+
+@asynccontextmanager
+async def lifespan(app: FastAPI):
+    """Startup and shutdown events."""
+    logger.info("🚀 BantayCam starting up...")
+
+    # Init database
+    await init_db()
+    logger.info("✅ Database initialized")
+
+    # Load AI models
+    logger.info("Loading AI models (this may take a moment)...")
+    vehicle_detector.load()
+    face_service.load()
+    logger.info("✅ AI models loaded")
+
+    # Ensure snapshot storage exists
+    Path(settings.STORAGE_PATH).mkdir(parents=True, exist_ok=True)
+
+    yield
+
+    # Shutdown
+    logger.info("Shutting down streams...")
+    await StreamManager.get().stop_all()
+    logger.info("👋 BantayCam shut down")
+
+
+app = FastAPI(
+    title="BantayCam",
+    description="Integrated License Plate & Face Recognition for Philippine CCTV Infrastructure",
+    version=settings.APP_VERSION,
+    lifespan=lifespan,
+)
+
+app.add_middleware(
+    CORSMiddleware,
+    allow_origins=["*"],  # Tighten in production
+    allow_credentials=True,
+    allow_methods=["*"],
+    allow_headers=["*"],
+)
+
+# API routes
+app.include_router(cameras.router, prefix="/api/v1")
+app.include_router(vehicles.router, prefix="/api/v1")
+app.include_router(persons.router, prefix="/api/v1")
+
+# Serve snapshots as static files
+snapshot_dir = Path(settings.STORAGE_PATH)
+snapshot_dir.mkdir(parents=True, exist_ok=True)
+app.mount("/snapshots", StaticFiles(directory=str(snapshot_dir)), name="snapshots")
+
+
+@app.get("/")
+async def root():
+    return {
+        "app": settings.APP_NAME,
+        "version": settings.APP_VERSION,
+        "status": "running",
+        "docs": "/docs",
+    }
+
+
+@app.get("/health")
+async def health():
+    return {"status": "ok"}
--- a/backend/app/models/init.py
+++ b/backend/app/models/init.py
--- a/backend/app/models/camera.py
+++ b/backend/app/models/camera.py
@@ -0,0 +1,30 @@
+from sqlalchemy import Column, String, Boolean, DateTime, Text, Enum
+from sqlalchemy.dialects.postgresql import UUID
+from sqlalchemy.sql import func
+import uuid
+import enum
+
+from app.core.database import Base
+
+
+class CameraStatus(str, enum.Enum):
+    active = "active"
+    inactive = "inactive"
+    error = "error"
+
+
+class Camera(Base):
+    __tablename__ = "cameras"
+
+    id = Column(UUID(as_uuid=True), primary_key=True, default=uuid.uuid4)
+    name = Column(String(255), nullable=False, unique=True)           # e.g. "Gate 1 - Main Entrance"
+    rtsp_url = Column(Text, nullable=False)                           # rtsp://admin:pass@192.168.1.64:554/...
+    location = Column(String(255), nullable=True)                     # e.g. "North Gate, Building A"
+    description = Column(Text, nullable=True)
+    status = Column(Enum(CameraStatus), default=CameraStatus.inactive)
+    is_enabled = Column(Boolean, default=True)
+    created_at = Column(DateTime(timezone=True), server_default=func.now())
+    updated_at = Column(DateTime(timezone=True), onupdate=func.now())
+
+    def __repr__(self):
+        return f"<Camera {self.name} ({self.status})>"
--- a/backend/app/models/person.py
+++ b/backend/app/models/person.py
@@ -0,0 +1,65 @@
+from sqlalchemy import Column, String, Integer, Float, DateTime, Text, ForeignKey, JSON
+from sqlalchemy.dialects.postgresql import UUID, ARRAY
+from sqlalchemy.sql import func
+from pgvector.sqlalchemy import Vector
+import uuid
+
+from app.core.database import Base
+
+
+class PersonIdentity(Base):
+    """
+    Unique person identity — grouped by face embedding similarity.
+    All face events linked to this identity let us trace where this person has been.
+    """
+    __tablename__ = "person_identities"
+
+    id = Column(UUID(as_uuid=True), primary_key=True, default=uuid.uuid4)
+    name = Column(String(255), nullable=True)                # e.g. "Juan Dela Cruz" if registered
+    label = Column(String(100), nullable=True)               # e.g. "Unit 3A Resident", "Delivery Rider"
+    notes = Column(Text, nullable=True)
+    is_registered = Column(Integer, default=0)               # 1 = manually registered with name
+    is_watchlisted = Column(Integer, default=0)              # 1 = flagged for alerts
+    thumbnail_path = Column(Text, nullable=True)             # Best face shot
+
+    # Face embedding (512-dim for InsightFace buffalo_l)
+    embedding = Column(Vector(512), nullable=True)
+
+    # Stats
+    first_seen_at = Column(DateTime(timezone=True), nullable=True)
+    last_seen_at = Column(DateTime(timezone=True), nullable=True)
+    total_sightings = Column(Integer, default=0)
+    created_at = Column(DateTime(timezone=True), server_default=func.now())
+
+    def __repr__(self):
+        return f"<PersonIdentity {self.name or 'Unknown'} ({self.total_sightings} sightings)>"
+
+
+class FaceEvent(Base):
+    """
+    Every face detection event — one row per detected face per frame.
+    """
+    __tablename__ = "face_events"
+
+    id = Column(UUID(as_uuid=True), primary_key=True, default=uuid.uuid4)
+    camera_id = Column(UUID(as_uuid=True), ForeignKey("cameras.id"), nullable=False, index=True)
+    identity_id = Column(UUID(as_uuid=True), ForeignKey("person_identities.id"), nullable=True, index=True)
+
+    # Detection
+    detection_confidence = Column(Float, nullable=True)
+    face_bbox = Column(JSON, nullable=True)                  # {"x","y","w","h"}
+    face_embedding = Column(Vector(512), nullable=True)      # Per-event embedding for re-clustering
+    similarity_score = Column(Float, nullable=True)          # Match score to identity
+
+    # Context — if detected alongside a vehicle event
+    vehicle_event_id = Column(UUID(as_uuid=True), ForeignKey("vehicle_events.id"), nullable=True)
+
+    # Storage
+    snapshot_path = Column(Text, nullable=True)             # Full frame
+    face_snapshot_path = Column(Text, nullable=True)        # Cropped face chip
+
+    # Meta
+    captured_at = Column(DateTime(timezone=True), server_default=func.now(), index=True)
+
+    def __repr__(self):
+        return f"<FaceEvent identity={self.identity_id} cam={self.camera_id} at={self.captured_at}>"
--- a/backend/app/models/vehicle.py
+++ b/backend/app/models/vehicle.py
@@ -0,0 +1,73 @@
+from sqlalchemy import Column, String, Integer, Float, DateTime, Text, ForeignKey, Enum, JSON
+from sqlalchemy.dialects.postgresql import UUID
+from sqlalchemy.sql import func
+import uuid
+import enum
+
+from app.core.database import Base
+
+
+class VehicleType(str, enum.Enum):
+    car = "car"
+    motorcycle = "motorcycle"
+    truck = "truck"
+    van = "van"
+    jeepney = "jeepney"
+    tricycle = "tricycle"
+    unknown = "unknown"
+
+
+class VehicleIdentity(Base):
+    """
+    Unique vehicle identity — grouped by license plate.
+    Tracks every time this vehicle was seen across any camera.
+    """
+    __tablename__ = "vehicle_identities"
+
+    id = Column(UUID(as_uuid=True), primary_key=True, default=uuid.uuid4)
+    plate_number = Column(String(20), nullable=True, index=True, unique=True)
+    vehicle_type = Column(Enum(VehicleType), default=VehicleType.unknown)
+    color = Column(String(50), nullable=True)
+    make_model = Column(String(100), nullable=True)          # e.g. "Toyota Vios"
+    notes = Column(Text, nullable=True)
+    label = Column(String(100), nullable=True)               # e.g. "Unit 4B Owner", "Delivery Van"
+    is_whitelisted = Column(Integer, default=0)              # 1 = resident/approved
+    is_blacklisted = Column(Integer, default=0)              # 1 = flagged/banned
+    thumbnail_path = Column(Text, nullable=True)             # Best shot saved
+    first_seen_at = Column(DateTime(timezone=True), nullable=True)
+    last_seen_at = Column(DateTime(timezone=True), nullable=True)
+    total_sightings = Column(Integer, default=0)
+    created_at = Column(DateTime(timezone=True), server_default=func.now())
+
+
+class VehicleEvent(Base):
+    """
+    Every detection event — one row per camera capture.
+    """
+    __tablename__ = "vehicle_events"
+
+    id = Column(UUID(as_uuid=True), primary_key=True, default=uuid.uuid4)
+    camera_id = Column(UUID(as_uuid=True), ForeignKey("cameras.id"), nullable=False, index=True)
+    identity_id = Column(UUID(as_uuid=True), ForeignKey("vehicle_identities.id"), nullable=True, index=True)
+
+    # Detection data
+    plate_number = Column(String(20), nullable=True)
+    plate_confidence = Column(Float, nullable=True)
+    vehicle_type = Column(Enum(VehicleType), default=VehicleType.unknown)
+    color = Column(String(50), nullable=True)
+    person_count = Column(Integer, nullable=True)            # For motorcycle: how many riders
+    detection_confidence = Column(Float, nullable=True)
+
+    # Bounding boxes (stored as JSON: {"x":0,"y":0,"w":100,"h":100})
+    vehicle_bbox = Column(JSON, nullable=True)
+    plate_bbox = Column(JSON, nullable=True)
+
+    # Storage
+    snapshot_path = Column(Text, nullable=True)             # Full frame snapshot
+    plate_snapshot_path = Column(Text, nullable=True)       # Cropped plate image
+
+    # Meta
+    captured_at = Column(DateTime(timezone=True), server_default=func.now(), index=True)
+
+    def __repr__(self):
+        return f"<VehicleEvent plate={self.plate_number} type={self.vehicle_type} at={self.captured_at}>"
--- a/backend/app/services/init.py
+++ b/backend/app/services/init.py
--- a/backend/app/services/face_service.py
+++ b/backend/app/services/face_service.py
@@ -0,0 +1,88 @@
+"""
+FaceService — InsightFace-based face detection, embedding, and identity matching.
+Uses pgvector cosine similarity for fast 1:N identity search.
+"""
+import cv2
+import numpy as np
+import logging
+from typing import Optional
+from dataclasses import dataclass, field
+from uuid import UUID
+
+from app.core.config import settings
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class FaceDetection:
+    bbox: dict = field(default_factory=dict)            # {"x","y","w","h"}
+    embedding: Optional[np.ndarray] = None              # 512-dim vector
+    detection_confidence: float = 0.0
+    face_crop: Optional[np.ndarray] = None
+    matched_identity_id: Optional[UUID] = None
+    similarity_score: float = 0.0
+
+
+class FaceService:
+    def __init__(self):
+        self._app = None
+        self._ready = False
+
+    def load(self):
+        """Load InsightFace model — call once at startup."""
+        try:
+            from insightface.app import FaceAnalysis
+            self._app = FaceAnalysis(
+                name=settings.FACE_MODEL,
+                allowed_modules=["detection", "recognition"],
+            )
+            self._app.prepare(
+                ctx_id=settings.GPU_DEVICE,
+                det_size=(640, 640),
+            )
+            self._ready = True
+            logger.info(f"InsightFace loaded (model={settings.FACE_MODEL})")
+        except Exception as e:
+            logger.error(f"Failed to load InsightFace: {e}")
+
+    def detect(self, frame: np.ndarray) -> list[FaceDetection]:
+        """
+        Detect all faces in a frame and extract embeddings.
+        Returns list of FaceDetection objects.
+        """
+        if not self._ready or self._app is None:
+            return []
+
+        try:
+            faces = self._app.get(frame)
+        except Exception as e:
+            logger.debug(f"Face detection error: {e}")
+            return []
+
+        detections = []
+        for face in faces:
+            bbox = face.bbox.astype(int)
+            x1, y1, x2, y2 = bbox[0], bbox[1], bbox[2], bbox[3]
+
+            det = FaceDetection(
+                bbox={"x": x1, "y": y1, "w": x2 - x1, "h": y2 - y1},
+                embedding=face.normed_embedding,  # Already L2-normalized
+                detection_confidence=float(face.det_score),
+                face_crop=frame[max(0, y1):y2, max(0, x1):x2],
+            )
+            detections.append(det)
+
+        return detections
+
+    def cosine_similarity(self, emb1: np.ndarray, emb2: np.ndarray) -> float:
+        """Cosine similarity between two normalized embeddings."""
+        return float(np.dot(emb1, emb2))
+
+    def embeddings_to_list(self, embedding: np.ndarray) -> list[float]:
+        """Convert numpy embedding to list for pgvector storage."""
+        return embedding.tolist()
+
+
+# Singleton
+face_service = FaceService()
--- a/backend/app/services/stream_manager.py
+++ b/backend/app/services/stream_manager.py
@@ -0,0 +1,153 @@
+"""
+StreamManager — Manages RTSP stream connections and dispatches frames to workers.
+Each camera runs in its own asyncio task with a frame queue.
+"""
+import asyncio
+import cv2
+import logging
+from typing import Dict, Optional
+from uuid import UUID
+from datetime import datetime
+
+from app.core.config import settings
+
+logger = logging.getLogger(__name__)
+
+
+class StreamWorker:
+    """
+    Handles one RTSP camera stream.
+    Reads frames and puts them into a queue for the AI pipeline to consume.
+    """
+
+    def __init__(self, camera_id: UUID, camera_name: str, rtsp_url: str):
+        self.camera_id = camera_id
+        self.camera_name = camera_name
+        self.rtsp_url = rtsp_url
+        self.frame_queue: asyncio.Queue = asyncio.Queue(maxsize=10)
+        self.is_running = False
+        self._task: Optional[asyncio.Task] = None
+        self.frame_count = 0
+        self.last_frame_at: Optional[datetime] = None
+        self.error: Optional[str] = None
+
+    async def start(self):
+        self.is_running = True
+        self._task = asyncio.create_task(self._read_loop())
+        logger.info(f"[{self.camera_name}] Stream started")
+
+    async def stop(self):
+        self.is_running = False
+        if self._task:
+            self._task.cancel()
+            try:
+                await self._task
+            except asyncio.CancelledError:
+                pass
+        logger.info(f"[{self.camera_name}] Stream stopped")
+
+    async def _read_loop(self):
+        """Read frames from RTSP in a thread pool (OpenCV is blocking)."""
+        loop = asyncio.get_event_loop()
+
+        while self.is_running:
+            try:
+                cap = await loop.run_in_executor(
+                    None,
+                    lambda: cv2.VideoCapture(self.rtsp_url)
+                )
+
+                if not cap.isOpened():
+                    self.error = "Cannot open RTSP stream"
+                    logger.error(f"[{self.camera_name}] {self.error}")
+                    await asyncio.sleep(5)
+                    continue
+
+                self.error = None
+                frame_idx = 0
+
+                while self.is_running:
+                    ret, frame = await loop.run_in_executor(None, cap.read)
+
+                    if not ret:
+                        logger.warning(f"[{self.camera_name}] Frame read failed, reconnecting...")
+                        break
+
+                    frame_idx += 1
+                    self.frame_count += 1
+                    self.last_frame_at = datetime.utcnow()
+
+                    # Skip frames for performance
+                    if frame_idx % settings.PROCESS_EVERY_N_FRAMES != 0:
+                        continue
+
+                    # Non-blocking put — drop frame if queue is full
+                    try:
+                        self.frame_queue.put_nowait({
+                            "frame": frame,
+                            "frame_idx": frame_idx,
+                            "captured_at": self.last_frame_at,
+                        })
+                    except asyncio.QueueFull:
+                        pass  # AI pipeline is slow — drop frame
+
+                cap.release()
+
+            except asyncio.CancelledError:
+                raise
+            except Exception as e:
+                self.error = str(e)
+                logger.exception(f"[{self.camera_name}] Stream error: {e}")
+                await asyncio.sleep(5)
+
+
+class StreamManager:
+    """
+    Singleton that manages all active camera streams.
+    """
+    _instance = None
+
+    def __init__(self):
+        self._workers: Dict[str, StreamWorker] = {}
+
+    @classmethod
+    def get(cls) -> "StreamManager":
+        if cls._instance is None:
+            cls._instance = StreamManager()
+        return cls._instance
+
+    async def start_stream(self, camera_id: UUID, camera_name: str, rtsp_url: str):
+        key = str(camera_id)
+        if key in self._workers:
+            await self.stop_stream(camera_id)
+
+        worker = StreamWorker(camera_id, camera_name, rtsp_url)
+        self._workers[key] = worker
+        await worker.start()
+        return worker
+
+    async def stop_stream(self, camera_id: UUID):
+        key = str(camera_id)
+        if key in self._workers:
+            await self._workers[key].stop()
+            del self._workers[key]
+
+    def get_worker(self, camera_id: UUID) -> Optional[StreamWorker]:
+        return self._workers.get(str(camera_id))
+
+    def get_all_status(self) -> list:
+        return [
+            {
+                "camera_id": str(k),
+                "is_running": w.is_running,
+                "frame_count": w.frame_count,
+                "last_frame_at": w.last_frame_at,
+                "error": w.error,
+            }
+            for k, w in self._workers.items()
+        ]
+
+    async def stop_all(self):
+        for worker in list(self._workers.values()):
+            await worker.stop()
+        self._workers.clear()
--- a/backend/app/services/vehicle_detector.py
+++ b/backend/app/services/vehicle_detector.py
@@ -0,0 +1,227 @@
+"""
+VehicleDetector — YOLOv8-based vehicle detection, classification, and person counting.
+Also handles license plate OCR via fast-alpr.
+"""
+import cv2
+import numpy as np
+import logging
+from pathlib import Path
+from typing import Optional
+from dataclasses import dataclass, field
+
+from app.core.config import settings
+
+logger = logging.getLogger(__name__)
+
+VEHICLE_CLASSES = {
+    2: "car",
+    3: "motorcycle",
+    5: "bus",
+    7: "truck",
+}
+
+# COCO person class
+PERSON_CLASS = 0
+
+# Color detection ranges (HSV)
+COLOR_RANGES = {
+    "red":    [(0, 70, 50), (10, 255, 255)],
+    "red2":   [(170, 70, 50), (180, 255, 255)],
+    "blue":   [(100, 70, 50), (130, 255, 255)],
+    "white":  [(0, 0, 180), (180, 30, 255)],
+    "black":  [(0, 0, 0), (180, 255, 50)],
+    "silver": [(0, 0, 100), (180, 30, 180)],
+    "yellow": [(20, 70, 50), (35, 255, 255)],
+    "green":  [(35, 70, 50), (85, 255, 255)],
+    "orange": [(10, 70, 50), (20, 255, 255)],
+}
+
+
+@dataclass
+class VehicleDetection:
+    vehicle_type: str = "unknown"
+    color: Optional[str] = None
+    plate_number: Optional[str] = None
+    plate_confidence: float = 0.0
+    person_count: int = 1
+    detection_confidence: float = 0.0
+    vehicle_bbox: dict = field(default_factory=dict)
+    plate_bbox: dict = field(default_factory=dict)
+    vehicle_crop: Optional[np.ndarray] = None
+    plate_crop: Optional[np.ndarray] = None
+
+
+class VehicleDetector:
+    def __init__(self):
+        self._yolo = None
+        self._alpr = None
+        self._ready = False
+
+    def load(self):
+        """Load models — call once at startup."""
+        try:
+            from ultralytics import YOLO
+            self._yolo = YOLO(settings.YOLO_MODEL)
+            logger.info("YOLO model loaded")
+        except Exception as e:
+            logger.error(f"Failed to load YOLO: {e}")
+            return
+
+        try:
+            from fast_alpr import ALPR
+            self._alpr = ALPR(
+                detector_model="plate-detection-v1-large",
+                ocr_model="global-plates-mobile-vit-v2-model",
+                device="cuda" if settings.GPU_DEVICE >= 0 else "cpu",
+            )
+            logger.info("ALPR model loaded")
+        except Exception as e:
+            logger.warning(f"ALPR not loaded (will skip plate reading): {e}")
+
+        self._ready = True
+
+    def detect(self, frame: np.ndarray) -> list[VehicleDetection]:
+        """
+        Run vehicle detection on a frame.
+        Returns list of VehicleDetection objects.
+        """
+        if not self._ready or self._yolo is None:
+            return []
+
+        results = self._yolo(
+            frame,
+            conf=settings.DETECTION_CONFIDENCE_THRESHOLD,
+            verbose=False,
+        )
+
+        detections = []
+        h, w = frame.shape[:2]
+
+        for result in results:
+            boxes = result.boxes
+            if boxes is None:
+                continue
+
+            # Group: find all vehicles and all persons
+            vehicles = []
+            persons = []
+
+            for box in boxes:
+                cls_id = int(box.cls[0])
+                conf = float(box.conf[0])
+                x1, y1, x2, y2 = map(int, box.xyxy[0])
+
+                if cls_id in VEHICLE_CLASSES:
+                    vehicles.append({
+                        "type": VEHICLE_CLASSES[cls_id],
+                        "conf": conf,
+                        "bbox": {"x": x1, "y": y1, "w": x2 - x1, "h": y2 - y1},
+                        "crop": frame[y1:y2, x1:x2],
+                    })
+                elif cls_id == PERSON_CLASS:
+                    persons.append({"bbox": {"x": x1, "y": y1, "w": x2 - x1, "h": y2 - y1}})
+
+            for vehicle in vehicles:
+                det = VehicleDetection(
+                    vehicle_type=vehicle["type"],
+                    detection_confidence=vehicle["conf"],
+                    vehicle_bbox=vehicle["bbox"],
+                    vehicle_crop=vehicle["crop"],
+                )
+
+                # Detect color
+                det.color = self._detect_color(vehicle["crop"])
+
+                # Count persons on/near motorcycle
+                if vehicle["type"] == "motorcycle":
+                    det.person_count = self._count_persons_on_motorcycle(
+                        vehicle["bbox"], persons
+                    )
+
+                # Read plate
+                if self._alpr is not None:
+                    plate_result = self._read_plate(frame, vehicle["bbox"])
+                    if plate_result:
+                        det.plate_number = plate_result["text"]
+                        det.plate_confidence = plate_result["confidence"]
+                        det.plate_bbox = plate_result["bbox"]
+                        bx = plate_result["bbox"]
+                        det.plate_crop = frame[
+                            bx["y"]:bx["y"] + bx["h"],
+                            bx["x"]:bx["x"] + bx["w"]
+                        ]
+
+                detections.append(det)
+
+        return detections
+
+    def _read_plate(self, frame: np.ndarray, vehicle_bbox: dict) -> Optional[dict]:
+        """Run ALPR on vehicle region."""
+        if self._alpr is None:
+            return None
+        try:
+            # Expand vehicle crop slightly for plate detection
+            x, y, w, h = vehicle_bbox["x"], vehicle_bbox["y"], vehicle_bbox["w"], vehicle_bbox["h"]
+            crop = frame[max(0, y):y + h, max(0, x):x + w]
+            results = self._alpr.run(crop)
+            if results:
+                best = results[0]
+                ocr = best.ocr
+                if ocr and ocr.text:
+                    # Offset bbox back to full frame coordinates
+                    pb = best.detection.bounding_box
+                    return {
+                        "text": ocr.text.upper().replace(" ", ""),
+                        "confidence": float(ocr.confidence),
+                        "bbox": {
+                            "x": x + int(pb.x1),
+                            "y": y + int(pb.y1),
+                            "w": int(pb.x2 - pb.x1),
+                            "h": int(pb.y2 - pb.y1),
+                        }
+                    }
+        except Exception as e:
+            logger.debug(f"ALPR error: {e}")
+        return None
+
+    def _detect_color(self, crop: np.ndarray) -> str:
+        """Detect dominant vehicle color using HSV histogram."""
+        if crop is None or crop.size == 0:
+            return "unknown"
+        try:
+            hsv = cv2.cvtColor(crop, cv2.COLOR_BGR2HSV)
+            max_pixels = 0
+            detected_color = "unknown"
+
+            for color_name, (lower, upper) in COLOR_RANGES.items():
+                mask = cv2.inRange(hsv, np.array(lower), np.array(upper))
+                pixel_count = cv2.countNonZero(mask)
+                if pixel_count > max_pixels:
+                    max_pixels = pixel_count
+                    detected_color = color_name.replace("2", "")  # red2 → red
+
+            return detected_color
+        except Exception:
+            return "unknown"
+
+    def _count_persons_on_motorcycle(self, moto_bbox: dict, persons: list) -> int:
+        """Count persons whose center falls within or near the motorcycle bbox."""
+        count = 0
+        mx, my, mw, mh = moto_bbox["x"], moto_bbox["y"], moto_bbox["w"], moto_bbox["h"]
+
+        for person in persons:
+            px, py, pw, ph = person["bbox"]["x"], person["bbox"]["y"], person["bbox"]["w"], person["bbox"]["h"]
+            # Person center
+            cx = px + pw // 2
+            cy = py + ph // 2
+            # Check if center is within motorcycle bounding box (with some margin)
+            margin = 30
+            if (mx - margin <= cx <= mx + mw + margin and
+                    my - margin <= cy <= my + mh + margin):
+                count += 1
+
+        return max(count, 1)  # At least 1 rider assumed
+
+
+# Singleton
+vehicle_detector = VehicleDetector()
--- a/backend/app/workers/init.py
+++ b/backend/app/workers/init.py
--- a/backend/app/workers/pipeline.py
+++ b/backend/app/workers/pipeline.py
@@ -0,0 +1,250 @@
+"""
+Detection Pipeline Worker
+Pulls frames from a StreamWorker queue and runs:
+  1. Vehicle detection (type, color, plate, person count)
+  2. Face detection (embedding, identity match)
+  3. Saves snapshots and events to DB
+"""
+import asyncio
+import cv2
+import logging
+import os
+from datetime import datetime, timezone
+from pathlib import Path
+from uuid import UUID
+
+from sqlalchemy.ext.asyncio import AsyncSession
+from sqlalchemy import select, text
+
+from app.core.config import settings
+from app.core.database import AsyncSessionLocal
+from app.models.camera import Camera
+from app.models.vehicle import VehicleEvent, VehicleIdentity, VehicleType
+from app.models.person import FaceEvent, PersonIdentity
+from app.services.vehicle_detector import vehicle_detector
+from app.services.face_service import face_service
+from app.services.stream_manager import StreamWorker
+
+logger = logging.getLogger(__name__)
+
+
+class DetectionPipeline:
+    """
+    Runs detection on frames for a single camera stream.
+    """
+
+    def __init__(self, camera_id: UUID, camera_name: str, worker: StreamWorker):
+        self.camera_id = camera_id
+        self.camera_name = camera_name
+        self.worker = worker
+        self.snapshot_dir = Path(settings.STORAGE_PATH) / str(camera_id)
+        self.snapshot_dir.mkdir(parents=True, exist_ok=True)
+        self._running = False
+        self._task = None
+
+    async def start(self):
+        self._running = True
+        self._task = asyncio.create_task(self._loop())
+        logger.info(f"[{self.camera_name}] Pipeline started")
+
+    async def stop(self):
+        self._running = False
+        if self._task:
+            self._task.cancel()
+            try:
+                await self._task
+            except asyncio.CancelledError:
+                pass
+
+    async def _loop(self):
+        loop = asyncio.get_event_loop()
+
+        while self._running:
+            try:
+                frame_data = await asyncio.wait_for(
+                    self.worker.frame_queue.get(), timeout=2.0
+                )
+            except asyncio.TimeoutError:
+                continue
+            except asyncio.CancelledError:
+                raise
+
+            frame = frame_data["frame"]
+            captured_at = frame_data["captured_at"]
+
+            try:
+                await loop.run_in_executor(
+                    None, self._process_frame, frame, captured_at
+                )
+            except Exception as e:
+                logger.exception(f"[{self.camera_name}] Pipeline error: {e}")
+
+    def _process_frame(self, frame, captured_at: datetime):
+        """Synchronous processing — runs in thread pool."""
+        import asyncio
+        loop = asyncio.new_event_loop()
+        asyncio.set_event_loop(loop)
+        try:
+            loop.run_until_complete(self._async_process(frame, captured_at))
+        finally:
+            loop.close()
+
+    async def _async_process(self, frame, captured_at: datetime):
+        """Run detections and save to DB."""
+        # --- Vehicle Detection ---
+        vehicle_detections = vehicle_detector.detect(frame)
+
+        # --- Face Detection ---
+        face_detections = face_service.detect(frame)
+
+        if not vehicle_detections and not face_detections:
+            return
+
+        # Save snapshot
+        ts = captured_at.strftime("%Y%m%d_%H%M%S_%f")
+        snapshot_filename = f"{ts}.jpg"
+        snapshot_path = str(self.snapshot_dir / snapshot_filename)
+        cv2.imwrite(snapshot_path, frame)
+
+        async with AsyncSessionLocal() as db:
+            vehicle_event_ids = []
+
+            # Save vehicle events
+            for det in vehicle_detections:
+                identity_id = await self._get_or_create_vehicle_identity(
+                    db, det.plate_number, det.vehicle_type, det.color
+                )
+
+                # Save plate crop
+                plate_snap = None
+                if det.plate_crop is not None and det.plate_crop.size > 0:
+                    plate_snap = str(self.snapshot_dir / f"{ts}_plate.jpg")
+                    cv2.imwrite(plate_snap, det.plate_crop)
+
+                event = VehicleEvent(
+                    camera_id=self.camera_id,
+                    identity_id=identity_id,
+                    plate_number=det.plate_number,
+                    plate_confidence=det.plate_confidence,
+                    vehicle_type=det.vehicle_type,
+                    color=det.color,
+                    person_count=det.person_count,
+                    detection_confidence=det.detection_confidence,
+                    vehicle_bbox=det.vehicle_bbox,
+                    plate_bbox=det.plate_bbox,
+                    snapshot_path=snapshot_path,
+                    plate_snapshot_path=plate_snap,
+                    captured_at=captured_at,
+                )
+                db.add(event)
+                await db.flush()
+                vehicle_event_ids.append(event.id)
+
+            # Save face events
+            for i, det in enumerate(face_detections):
+                identity_id, similarity = await self._match_or_create_person(
+                    db, det.embedding
+                )
+
+                # Save face crop
+                face_snap = None
+                if det.face_crop is not None and det.face_crop.size > 0:
+                    face_snap = str(self.snapshot_dir / f"{ts}_face{i}.jpg")
+                    cv2.imwrite(face_snap, det.face_crop)
+
+                embedding_list = face_service.embeddings_to_list(det.embedding) if det.embedding is not None else None
+
+                event = FaceEvent(
+                    camera_id=self.camera_id,
+                    identity_id=identity_id,
+                    detection_confidence=det.detection_confidence,
+                    face_bbox=det.bbox,
+                    face_embedding=embedding_list,
+                    similarity_score=similarity,
+                    vehicle_event_id=vehicle_event_ids[0] if vehicle_event_ids else None,
+                    snapshot_path=snapshot_path,
+                    face_snapshot_path=face_snap,
+                    captured_at=captured_at,
+                )
+                db.add(event)
+
+            await db.commit()
+
+    async def _get_or_create_vehicle_identity(
+        self, db: AsyncSession, plate_number, vehicle_type, color
+    ):
+        """Find existing vehicle identity by plate or create new one."""
+        if not plate_number:
+            return None
+
+        result = await db.execute(
+            select(VehicleIdentity).where(VehicleIdentity.plate_number == plate_number)
+        )
+        identity = result.scalar_one_or_none()
+
+        if identity:
+            identity.last_seen_at = datetime.now(timezone.utc)
+            identity.total_sightings += 1
+            if color and not identity.color:
+                identity.color = color
+        else:
+            identity = VehicleIdentity(
+                plate_number=plate_number,
+                vehicle_type=vehicle_type or VehicleType.unknown,
+                color=color,
+                first_seen_at=datetime.now(timezone.utc),
+                last_seen_at=datetime.now(timezone.utc),
+                total_sightings=1,
+            )
+            db.add(identity)
+            await db.flush()
+
+        return identity.id
+
+    async def _match_or_create_person(self, db: AsyncSession, embedding):
+        """
+        Find closest matching person identity using pgvector cosine similarity.
+        If no match above threshold, create new identity.
+        Returns (identity_id, similarity_score).
+        """
+        if embedding is None:
+            return None, 0.0
+
+        emb_list = face_service.embeddings_to_list(embedding)
+        emb_str = f"[{','.join(str(x) for x in emb_list)}]"
+
+        # pgvector cosine distance (1 - cosine_similarity)
+        threshold = settings.FACE_SIMILARITY_THRESHOLD  # distance threshold
+
+        result = await db.execute(
+            text(f"""
+                SELECT id, 1 - (embedding <=> '{emb_str}'::vector) AS similarity
+                FROM person_identities
+                WHERE embedding IS NOT NULL
+                  AND (embedding <=> '{emb_str}'::vector) < :threshold
+                ORDER BY embedding <=> '{emb_str}'::vector
+                LIMIT 1
+            """),
+            {"threshold": threshold}
+        )
+        row = result.fetchone()
+
+        if row:
+            identity_id, similarity = row
+            # Update last seen
+            await db.execute(
+                text("UPDATE person_identities SET last_seen_at=NOW(), total_sightings=total_sightings+1 WHERE id=:id"),
+                {"id": identity_id}
+            )
+            return identity_id, float(similarity)
+        else:
+            # New unknown person
+            identity = PersonIdentity(
+                first_seen_at=datetime.now(timezone.utc),
+                last_seen_at=datetime.now(timezone.utc),
+                total_sightings=1,
+                embedding=emb_list,
+            )
+            db.add(identity)
+            await db.flush()
+            return identity.id, 0.0
--- a/backend/requirements.txt
+++ b/backend/requirements.txt
@@ -0,0 +1,27 @@
+# Web framework
+fastapi==0.115.0
+uvicorn[standard]==0.30.6
+pydantic==2.9.2
+pydantic-settings==2.5.2
+
+# Database
+sqlalchemy[asyncio]==2.0.35
+asyncpg==0.29.0
+pgvector==0.3.2
+alembic==1.13.3
+
+# AI / Computer Vision
+ultralytics==8.3.0          # YOLOv8/v11 vehicle + person detection
+fast-alpr==0.2.0            # License plate recognition
+insightface==0.7.3          # Face detection + recognition (ArcFace)
+opencv-python-headless==4.10.0.84
+numpy==1.26.4
+onnxruntime-gpu==1.19.2     # GPU inference (use onnxruntime for CPU-only)
+
+# Optional: GPU acceleration
+# torch>=2.0 (install separately with CUDA support)
+
+# Utilities
+python-multipart==0.0.9     # File uploads
+aiofiles==24.1.0
+httpx==0.27.2
--- a/docker-compose.yml
+++ b/docker-compose.yml
@@ -0,0 +1,37 @@
+version: "3.9"
+
+services:
+  postgres:
+    image: pgvector/pgvector:pg16
+    container_name: bantaycam-db
+    environment:
+      POSTGRES_USER: bantaycam
+      POSTGRES_PASSWORD: bantaycam
+      POSTGRES_DB: bantaycam
+    ports:
+      - "5432:5432"
+    volumes:
+      - pgdata:/var/lib/postgresql/data
+    restart: unless-stopped
+
+  backend:
+    build:
+      context: ./backend
+      dockerfile: Dockerfile
+    container_name: bantaycam-api
+    depends_on:
+      - postgres
+    environment:
+      DATABASE_URL: postgresql+asyncpg://bantaycam:bantaycam@postgres:5432/bantaycam
+      STORAGE_PATH: /app/snapshots
+      GPU_DEVICE: -1          # Change to 0 for GPU
+    ports:
+      - "8000:8000"
+    volumes:
+      - ./snapshots:/app/snapshots
+      - ./backend:/app
+    restart: unless-stopped
+    command: uvicorn app.main:app --host 0.0.0.0 --port 8000 --reload
+
+volumes:
+  pgdata: