Computer vision · Capstone
Hybrid Self-Checkout
Barcode scanning paired with computer vision to verify retail items.
The problem
A barcode alone cannot tell whether the item in front of the camera is the item being scanned. Missed scans and item switching create gaps in conventional self-checkout.
What I built
I built a hybrid validation pipeline that identifies local retail products with YOLOv8 and compares detections with USB barcode scanner input. Python and OpenCV handle the video stream; MySQL stores product and transaction data. Split-view experiments address occlusion by observing products from multiple angles.
Constraints & lessons
Similar packaging, overlapping items, lighting changes, and missed detections make visual verification difficult. The model was trained on a custom dataset of Malaysian retail items, including bread, tissues, toothpaste, and Maggi. Raising recall remains an improvement target.
Result & evidence
The project repository reports approximately 77.4% precision and 72.0% recall after 50 epochs. These are model evaluation figures, not a measured reduction in retail fraud. The portfolio and resume disagree on inference latency, so no latency figure is presented here.
SOURCE NOTES
Project README & architecture ↗