AI-based machine vision for retail self-checkout system

Abstract: In recent years advances in computing power, availability of large annotated datasets and AI algorithms have enabled the rise of reliable object identification and tracking. This thesis describes the development and implementation considerations of a system for object detection in the retail industry. This project have been conducted as a collaboration between ETH Zürich and start-up company AI Retailing Systems, who wants to automate parts of the retailing experience, namely the checkout procedure in a retail store. A data set of images with corresponding bounding boxes and pixel-segmentations has been gathered, consisting of ten Swiss retail products. Relevant theory is discussed and three state of the art neural network architectures are reviewed and evaluated for the specific application and dataset. The thesis concludes with a discussion of the main challenges for this type of solution, a recommendation for the object detection model to be used and pointers for future work.