Real-time object detection using YOLO models : dataset creation, training and practical implementation

Abstract

Supermarket checkout is still a slow and labor-intensive process in many countries. Computer vision offers a promising way to automate it. However, most research on automatic product detection uses very large datasets that take significant time and resources to build, making it difficult to apply in practice. This thesis explores whether a reliable real-time product detection system can be built using a small, manually collected dataset of common everyday grocery items.

A custom dataset of 380 images across 10 packaged product classes was created and used to train a YOLOv11m deep learning model, also taking advantage of transfer learning and this way compensate for the limited amount of training data. The training setting and design choices were validated through running several experiments. The trained model correctly detected products with a mAP@0.5 of 0.990 and ran at 27 frames per second on a GPU, meeting real-time requirements. The augmentation experiment showed that training without augmentation led to a significant drop in accuracy and clear signs of overfitting, confirming that augmentation is critical when training data is limited.

The model was integrated into a working web application that uses a live camera feed to detect products, build a shopping cart and calculate a total price. The results show that accurate real-time retail product detection is achievable with a small custom dataset, though a production-ready system would need more training data and further optimization to handle more difficult real-world conditions reliably.