Retail CV Lite — Edge

Lightweight computer vision application for retail analytics on Jetson Orin Nano (and NVIDIA GPU workstations). Built on Savant + DeepStream 7.

Features

Feature	Details
People detection	NVIDIA PeopleNet (ResNet34, quantized ONNX)
Tracking	NvSORT (lightweight for edge)
Tripwire / line counting	Configurable in/out counts per line per camera
Zone dwell time	Min / avg / max dwell by hour and by day
Live streams	HLS, RTSP, WebRTC via Savant always-on-sink
Analytics storage	Elasticsearch (2 indices: counting + dwell)
Dashboards	Built-in React UI + Kibana for advanced charts
Deployment	Docker Compose for x86 and Jetson L4T

---

Architecture

RTSP Camera(s)
      │
      ▼ ZMQ (dealer+connect)
┌─────────────────────────────────┐
│        Savant Module            │
│  PeopleNet → Tracker →          │
│  LineCrossing pyfunc →          │  ──► Elasticsearch
│  ZoneDwell pyfunc               │       cv_lite_people_counting
└─────────────────────────────────┘       cv_lite_zone_dwell
      │ ZMQ (pub)
      ▼
Always-On Sink  ──► HLS :888 / RTSP :554 / WebRTC :8889
                          │
                          ▼
               React Frontend :3000
               (live stream + stats)

FastAPI Backend :8000
  ├── SQLite (camera/line/zone config)
  ├── writes line_config.yml & zone_config.yml → /config volume
  ├── REST API for CRUD
  └── Analytics API (queries Elasticsearch)

---

Quick Start (x86 / NVIDIA GPU)

# 1. Copy env file
cp .env.example .env
# Edit .env — set SITE_ID, ES_PASSWORD, etc.

# 2. Start core services
docker compose up -d

# 3. Open UI
open http://localhost:3000
# Kibana (optional)
open http://localhost:5601

Add cameras via the UI

1. Go to Configuration → Add Camera
2. Enter Camera ID (becomes Savant source_id), name, RTSP URL
3. Click Snapshot to preview the frame
4. Click Add Line → draw a tripwire on the frame → name it
5. Click Add Zone → click polygon points → double-click to close → name it

Start RTSP sources

For each camera start one source adapter container:

SOURCE_ID=cam-001 RTSP_URL=rtsp://user:pass@192.168.1.10:554/stream \
  docker compose --profile sources up rtsp-source -d

---

Quick Start (Jetson Orin Nano / L4T)

# Prerequisites: JetPack 6.x, Docker, nvidia-container-toolkit

cp .env.example .env
# Set ES_JAVA_OPTS=-Xms256m -Xmx256m in .env

docker compose -f docker-compose.l4t.yml up -d

# Kibana is optional — start with:
docker compose -f docker-compose.l4t.yml --profile kibana up kibana -d

First run note: TensorRT engine compilation for PeopleNet takes ~10 min on the first start. Subsequent starts are instant from cache.

---

Elasticsearch Indices

cv_lite_people_counting

Field	Type	Description
@timestamp	date	Event time
site_id	keyword	Site identifier
camera_id	keyword	Savant source ID
line_id	keyword	Tripwire line ID
direction	keyword	in or out
count_in	integer	Cumulative in count
count_out	integer	Cumulative out count

cv_lite_zone_dwell

Field	Type	Description
@timestamp	date	Exit time
zone_id	keyword	Zone identifier
track_id	long	Person track ID
enter_time	date	Zone entry timestamp
exit_time	date	Zone exit timestamp
dwell_seconds	float	Time in zone (s)

---

Optional: Person Attributes Model (YOLOv8)

To enable your YOLOv8 attributes model:

1. Set ENABLE_ATTRIBUTES_MODEL=true in .env
2. Mount your model weights into the savant-module container
3. Add a secondary nvinfer@classifier element in savant-module/module.yml

---

Config Files

The backend writes two YAML files to the shared /config volume that the Savant pyfuncs poll every 30 seconds:

• /config/line_config.yml — tripwire line coordinates per camera
• /config/zone_config.yml — zone polygon coordinates per camera

---

Project Structure

retail-cv-lite-edge/
├── savant-module/          # Savant/DeepStream pipeline
│   ├── module.yml          # Pipeline definition
│   ├── Dockerfile.x86
│   ├── Dockerfile.l4t
│   └── analytics/
│       ├── line_crossing.py
│       ├── zone_dwell.py
│       ├── config_loader.py
│       ├── es_client.py
│       └── overlay.py
├── backend/                # FastAPI REST API
│   ├── app/
│   │   ├── main.py
│   │   ├── models.py
│   │   ├── config_writer.py
│   │   └── routers/
│   └── Dockerfile
├── frontend/               # React + Tailwind UI
│   ├── src/
│   │   ├── pages/Dashboard.jsx
│   │   ├── pages/Configuration.jsx
│   │   └── components/CanvasEditor.jsx
│   └── Dockerfile
├── assets/tracker/         # NvSORT tracker config
├── docker-compose.yml      # x86 deployment
├── docker-compose.l4t.yml  # Jetson deployment
└── .env.example

---

Brand / UI Customisation

The UI uses a dark analytics theme (navy + sky-blue accents). To apply your brand colours, edit frontend/tailwind.config.js — update the brand colour palette.

Brand guidelines from the Google Slides presentation were not accessible at build time (requires Google auth). Share the colour hex codes and we can update the theme.