Empty Shelf Space Detection and Analytics

📝 Overview

This project presents a production-ready retail shelf analytics system that detects products and empty shelf spaces using unified YOLO11 models.

The system is designed for automated shelf monitoring in supermarkets and retail stores, enabling detection of:

• Product presence
• Empty shelf spaces
• Potential out-of-stock zones

This solution is relevant for large retail chains & smart stores, where manual shelf auditing is inefficient and error-prone.

---

⚠️ Problem Statement

Retail stores frequently suffer from:

• Out-of-stock products
• Poor shelf utilization
• Delayed restocking decisions

Manual shelf inspection does not scale across large supermarkets.
This project addresses this problem using computer vision-driven shelf analytics.

---

⚙️ Key Features

✔ Product detection on retail shelves
✔ Empty shelf space detection
✔ Unified dataset for products + empty shelf regions
✔ Per-row analytics for actionable insights
✔ Interval-based inference to save CPU/GPU resources
✔ CSV-based reporting for business dashboards
✔ Works with IP cameras, webcams, and video files
✔ Edge-device friendly (Jetson Nano, Raspberry Pi)

---

🗂️ Dataset Strategy (Key Innovation)

This project uses a hybrid dataset construction approach to enable robust empty shelf detection.

1️⃣ Product Detection Dataset

• SKU-based retail product images
• Trained YOLO11 model to detect products on shelves

2️⃣ Empty Shelf Dataset

• Public dataset from Roboflow:
  • Supermarket Empty Shelf Detector
• Contains only empty shelf annotations

3️⃣ Weakly-Supervised Relabeling (Core Contribution)

To unify both datasets: run save_label.py

1. Pretrained product YOLO11 model is run on empty shelf images
2. Detected products are auto-saved as YOLO labels
3. Final class mapping:
   • class_id = 0 → Empty shelf space
   • class_id = 1 → Product

Result: a single dataset with both classes without manual relabeling.

---

🏷️ Class Mapping

Class ID	Label
0	Empty Shelf Space
1	Product

---

🧩 System Architecture

The system processes video streams through three stages:

1. Product and Empty Space Detection (YOLO11)
   Detects all products and the empty space on the shelf.

2. Row Extraction (YOLO11 Seg)
   A YOLO segmentation model trained on shelf row annotations extracts row boundaries.

This enables accurate row-level shelf analytics.

---

⏱️ Interval-Based Inference (Resource Efficient)

Instead of processing every frame, the system:

• Reads camera continuously
• Runs inference once every N seconds

Benefits: ✔ Saves GPU/CPU resources
✔ Reduces power and thermal load
✔ Ideal for 24/7 edge deployments
✔ Works on low-bandwidth camera feeds

---

📊 Analytics Output (CSV → Dashboard Ready)

Each inference logs structured analytics into CSV, containing:

Timestamp	Row	Empty Slots	Empty %
2026-01-20 13:50:45	1	3	22.5%
2026-01-20 13:50:45	2	1	22.5%
2026-01-20 13:50:45	3	0	22.5%

The CSV output can be used for:

✔ Retail dashboards (Tableau / PowerBI / Superset)
✔ Automated alerts & notifications
✔ Inventory & ERP system integration

---

📦 Requirements

ultralytics
opencv-python
numpy
pandas

---

🏋️‍♂️ Training and Inference

Paste the below CLI to train

• for detection model

yolo detect train data=data.yaml model=yolo11n.pt epochs=100 imgsz=640

• for segmentation model

yolo segment train data=row_data.yaml model=yolo11n-seg.pt epochs=80 imgsz=640

Detection and Segmentation Results

(a) Detection	(b) Segmentation

For more hyperparameters see the Ultralytics documentation

• for inference

🎥 Works with multiple video sources:

✔ IP Cameras
✔ USB/Webcam
✔ Video Files (.mp4, .avi)
✔ Network Streams

python3 main.py --interval 30 --source "http://192.168.0.100:8080/video"

Real Time Inferenced

Detection Results and Shelf Statistics

(a) Detection Result	(b) Shelf Status Summary
	Timestamp-wise Shelf Analysis Timestamp Row Empty Slots Whole Empty Shelf % 2026-01-20 13:50:45 1 0 4.76% 2026-01-20 13:50:45 2 1 4.75% 2026-01-20 13:50:45 3 1 4.76% 2026-01-20 13:50:45 4 2 4.76% 2026-01-20 13:50:45 5 0 4.75% 2026-01-20 13:50:45 6 0 4.76% 2026-01-20 13:50:45 7 0 4.76%

---

📐 Empty Space Area Calculation

To quantify how much shelf space is empty vs occupied, the system uses an area-based metric instead of pure object count.

For each inference frame:

• All empty shelf regions (class_id = 0) are detected
• All product regions (class_id = 1) are detected

The area of each bounding box is computed and accumulated.

🧮 Formula

Empty Area % = (Total Empty Area / (Total Empty Area + Total Product Area)) * 100

Where:

• Total Empty Area = sum of pixel areas of all class 0 detections
• Total Product Area = sum of pixel areas of all class 1 detections

📝 Example Interpretation

If in a given frame:

• Total empty area = 120000 pixels
• Total product area = 480000 pixels

Then:

Empty Area % = (120000 / (120000 + 480000)) * 100 = (120000 / 600000) * 100 = 20%

Meaning 20% of the detected shelf area is empty.

🎯 Why Area-Based Metric?

Area-based measurement is preferred because:

✔ Items vary in size (e.g., small gum packs vs large cereal boxes)
✔ Shelf utilization is proportional to space, not just count
✔ More realistic for retail inventory analytics
✔ Works even when slot sizes are inconsistent

---

🚀 Deployment Workflow

Typical deployment flow:

• Mount camera facing shelf
• Start monitoring script with interval
• System logs per-row empty metrics
• CSV consumed by dashboards / APIs
• Alerts trigger when shelf empties

Optimized for edge devices & 24/7 store monitoring.

---

🛒 Applications

• Automated shelf monitoring
• Out-of-stock detection
• Shelf utilization analysis
• Retail analytics dashboards
• Smart store deployments

---

🌍 Beyond Retail — Other Use Cases

This method can also be applied to:

• Pharmacy shelves
• Warehouse bins
• Tool inventory racks
• Vending machines
• Library shelves

Where stock presence & empty space matter.

---

🔮 Future Enhancements

Planned improvements:

• Live dashboard UI (Flask/React)
• REST/MQTT alerting system
• Specific Product Alert
• Multi-camera store support
• Cloud & POS system integration
• Jetson / edge-accelerated builds

---