Retail Analytics App

Edge AI system for retail analytics — person tracking, cross-camera re-identification, and loitering detection. Runs on NVIDIA Jetson hardware using DeepStream and TensorRT.

---

Architecture

RTSP Cameras
    ↓
deepstream_service  — C++ DeepStream pipeline (detection, tracking, ReID)
    ↓ ZeroMQ
backend             — FastAPI (camera management, event storage, loitering rules)
    ↓
frontend            — Next.js dashboard
    ↓
nginx               — Reverse proxy (port 80)

See PLATFORMS.md for OpenVINO (x86) and RK3588 deployment variants.

---

Quick Start — Jetson

Prerequisites

• NVIDIA Jetson Orin NX / AGX / Xavier with JetPack 5.x
• Docker with nvidia runtime as default (docker info | grep "Default Runtime")
• NGC account with API key — ngc.nvidia.com

1. Clone

git clone --recurse-submodules git@github.com:atriva-ai/retail-analytics-app.git
cd retail-analytics-app

2. Download AI models (one-time per device)

Models are not stored in Git. Download them before building the Docker image. TRT engines are auto-generated from these ONNX files on first container start and written back to the same directory.

cd services/deepstream-jetson
export NGC_API_KEY=<your_key>
./scripts/download_peoplenet.sh   # → models/peoplenet/resnet34_peoplenet_int8.onnx
./scripts/download_reid.sh        # → models/reid/resnet50_market1501_aicity156.onnx
cd ../..

3. Configure environment

cp backend/.env.example backend/.env
# Edit backend/.env to set POSTGRES_USER, POSTGRES_PASSWORD, POSTGRES_DB

4. Build and start

docker compose -f docker-compose.jetson.yml up --build

First run takes 8–12 minutes — Docker image builds (~5 min) then TRT engine generation (~3–5 min, one-time). Subsequent starts are ~30 seconds.

All 5 services are ready when you see:

deepstream_service  | ZMQ subscriber started
backend             | Application startup complete.
frontend            | ▲ Next.js ready

5. Register cameras and verify

# Register RTSP streams
curl -X POST "http://localhost:8001/cameras/register?camera_id=cam1&rtsp_url=rtsp://<host>:<port>/stream1"

# Check pipeline is live
curl http://localhost:8001/api/v1/video-pipeline/debug/
# → {"pipeline_alive": true, ...}

# Check detections are flowing
curl http://localhost:8001/shared/cameras/cam1/detections/latest

# Open dashboard
open http://localhost

For detailed DeepStream pipeline documentation, model configuration, and diagnostic tools, see services/deepstream-jetson/README.md.

---

Project Structure

retail-analytics-app/
├── frontend/                     # Next.js dashboard
├── backend/                      # FastAPI — camera registry, detection store, loitering API
├── services/
│   └── deepstream-jetson/        # C++ DeepStream pipeline + Python FastAPI shim (Jetson)
├── nginx/
│   ├── conf.d/                   # Shared nginx config (OpenVINO / RK3588)
│   └── conf.d.jetson/            # Jetson-specific nginx config
├── docker-compose.jetson.yml     # Jetson full-stack compose
├── PLATFORMS.md                  # Multi-platform build and deployment guide
└── docs/test-reports/            # Test reports

---

Stopping and restarting

# Stop (keeps volumes/data)
docker compose -f docker-compose.jetson.yml down

# Restart without rebuild (uses cached TRT engines — fast)
docker compose -f docker-compose.jetson.yml up

# Full rebuild (needed after code changes)
docker compose -f docker-compose.jetson.yml up --build

---

Environment variables

Variable	Service	Default	Purpose
POSTGRES_USER	db, backend	atriva	Database user
POSTGRES_PASSWORD	db, backend	—	Database password (set in backend/.env)
POSTGRES_DB	db, backend	atrv-retail	Database name
CAMERAS_JSON	deepstream_service	—	Pre-seed cameras: {"cam1":"rtsp://..."}
NGC_API_KEY	host (download scripts)	—	NGC model download