The situation
The store needed to protect stock from theft and keep inventory accurate without constant manual counts.
Custom retail item detection in store
Shelf-level detection that keeps stock counts current and flags missing or misplaced items for inspection.
A toy retailer
01 · Context
02 · Challenge
What had to be true
Recognise every toy on the shelf in real time and compare what the camera sees with the checkout system.
03 · What we built
The system
- A dataset of every retail item in different backgrounds and orientations, randomly resized and skewed
- labelImg annotations, an 80/20 split and a custom YAML file with all toy labels
- YOLOv5 trained to detect toys on the shelf
- Optimised from PyTorch to a TensorRT engine for the Jetson Nano GPU, running on the live stream
- Stock levels per location compared with the checkout inventory, with mismatches flagged for manual inspection
04 Architecture
How it works, step by step.
Step through the system, or let it play.
- Shelf camera streams video
- YOLOv5 detects each item
- Stock levels updated
- Compared with checkout system
- Mismatches flagged
05 · Outcome
Stock levels stay current from the camera, and possible theft or misplaced items are flagged quickly.
06 · Stack
Next case · Precious metals · Computer visionAutomated gold weight and purity detection