Machine vision techniques applied to the checkout process

Abstract

Traditionally, barcodes have been used to mark items in a machine readable format. This allowed an unknown item to be identified in a short time when compared to manual identification. More recently RFIDs have been proposed and started being used to identify unknown items. Both methods have their problems in limiting them to speed up and automate the item identification process in a supermarket checkout. Machine vision techniques, nowadays, are being applied to all sorts of applications ranging from consumer applications to industrial and scientific ones. This work proposes a possible method that uses machine vision to help in automating the checkout process in supermarkets. The main challenge to apply machine vision in this process is to be able to identify an unknown item in a very short time. In order to achieve this, an offline indexing algorithm to select a number of key object images through clustering was implemented. After the key images are selected, the distances between each key image and each object image are computed. The key image distances are then used to apply the triangle inequality principle during retrieval in order to reduce the number of object images to be directly compared. Furthermore, the length-width ratio of the unknown object is used so that objects which have a different ratio are immediately discarded. The system was tested with different clustering methods and distance measures and achieved a worst case accuracy of 84% when tested using the Amsterdam Library of Object Images. A more real life test using supermarket items and a specially built jig to provide constant lighting levels showed that the system reached an accuracy of 98%.