Computer Vision for Retail

A computer vision intelligence system for retail and mall environments — from model inference to business analytics.

Built as part of the Inside AI with Kayana YouTube channel. This project shows you how a standard camera network can be turned into an intelligent data system — one that feeds marketing, operations, and security teams with real, actionable insights.

---

What This Project Does

This system layers five computer vision technologies on top of a standard mall camera network:

Technology	What It Produces
Emotion Classification	Capture emotional state of shoppers
Multi-Camera ReID	Cross-camera identity matching and customer journey visualisation
Crowd Analysis	Zone-level occupancy counts and trend tracking
Intrusion Detection	Automated alerts when restricted zones are breached
BI Dashboard	Unified Streamlit dashboard connecting all data streams
Real-Time Notifications	Telegram alerts powered by Postgres LISTEN/NOTIFY

---

Demo

📹 Video walkthrough coming soon on Inside AI with Kayana

---

Tech Stack

Layer	Tool
People Detection	NVIDIA PeopleNet (ONNX)
Re-Identification	OSNet via torchreid
Emotion Classification	DeepFace
Zone Configuration	Roboflow PolygonZone
Database	PostgreSQL (local)
Dashboard	Streamlit
Notifications	Telegram Bot API + Postgres LISTEN/NOTIFY
Demo Environment	Jupyter Notebook (VS Code)
Language	Python 3.13

---

Project Structure

retail-deep-dive-computer-vision/
├── .streamlit/
├── models/
│   └── peoplenet.onnx
├── Input_Videos/
│   ├── Emotion_Classification
│   ├── ReID
│   ├── Crowd_Analysis
│   ├── Intrusion_Detection
├── Output_Videos/
│   ├── Emotion_Classification
│   ├── ReID
│   ├── Crowd_Analysis
│   ├── Intrusion_Detection
├── zone_setup_frames/
├── alert_crops/
├── synthetic_data_export/
├── demo_code.ipynb       ← main notebook (inference pipeline)
├── dashboard_combined_app.py            ← Streamlit analytics dashboard with real and synthetic data combined
├── dashboard_real_app.py            ← Streamlit analytics dashboard with real data only
├── notification_service.py        ← real-time Telegram alert service
└── README.md

Database

Five PostgreSQL tables store all CV outputs:

zones              ← polygon zone definitions per camera
rules              ← notification trigger rules
detection_events   ← all detection metadata from every use case
emotion_events     ← DeepFace results linked to detection_events
notifications      ← triggered alerts linked to rules and detections
reid_features      ← feature vectors for multicamera reidentification

---

Setup

1. Clone the repository

git clone https://github.com/elmayana/computer-vision-for-retail.git
cd computer-vision-for-retail

2. Create and activate a virtual environment

python -m venv venv
venv\Scripts\activate        # Windows
source venv/bin/activate     # macOS/Linux

3. Install dependencies and pytorch

install requirements.txt file here

pip install -r requirements.txt

pip install torch torchvision --index-url https://download.pytorch.org/whl/cuXXX 

where [XXX] is the CUDA version (check using nvidia-smi command)

4. Set up PostgreSQL

• Install PostgreSQL from https://www.postgresql.org/download/
• Create a database
• Update DB_PASSWORD and DB_USERNAME with your password and username in demo-code.ipynb
• in the .streamlit folder, create a secrets.toml file

[connections.postgresql]
dialect = "postgresql"
host = "localhost"
port = "your_port"
database = "your_database"
username = "your_username"
password = "your_password"

5. Download PeopleNet

• Download deployable_quantized_onnx_v2.6.3 from the PeopleNet NGC page
• Place it in the models/ folder

6. Download Input Videos

Section	Download
Emotion Classification	Emotion Video 1 Emotion Video 2
ReID (2 camera views)	Download ReID videos
Crowd Analysis	Pexels — Book Shop Time Lapse
Intrusion Detection	Pexels — Corridor Video

Place all videos in the Input_Videos/ folder and update the file paths

7. Set up Telegram Notifications

To receive real-time alerts on Telegram:

1. Message @BotFather on Telegram and create a new bot — you'll receive a bot token
2. Start a conversation with your new bot, then retrieve your chat ID by visiting:

   https://api.telegram.org/bot<YOUR_TOKEN>/getUpdates
   

3. Add both values to notification_service.py:

   TELEGRAM_BOT_TOKEN = "your_bot_token"
   TELEGRAM_CHAT_ID = "your_chat_id"

---

---

Main inference pipeline

Open the notebook in VS Code (or Jupyter Lab) and run the cells section by section.

Streamlit dashboard

Once the inference pipeline has run and your CSVs or database are populated:

streamlit run dashboard_combined_app.py

The dashboard opens in your browser at http://localhost:8501. It includes:

• Customer journey Sankey diagram (cross-zone flow)
• Emotion composition by zone and hour
• Floor plan heatmap
• Crowd and intrusion trend charts
• Sidebar date range filter

Real-time notification service

Run this in a separate terminal while the inference pipeline is active:

python notification_service.py

This service listens to Postgres for new events and sends a Telegram message whenever a configured rule is triggered. Security alerts (intrusion) and operational alerts (crowd thresholds) have separate debounce windows to prevent alert fatigue.

---

---

References

• NVIDIA PeopleNet on NGC
• NVIDIA TAO DetectNet_v2 Documentation
• DeepFace GitHub
• Roboflow PolygonZone
• Pexels — Royalty Free Retail Footage

---

License

This project is licensed under the GPL-3.0 License.