Computer Vision and Deep Learning for retail store management

Tonioni, Alessio (2019) Computer Vision and Deep Learning for retail store management, [Dissertation thesis], Alma Mater Studiorum Università di Bologna. Dottorato di ricerca in Computer science and engineering, 31 Ciclo. DOI 10.6092/unibo/amsdottorato/8970.
Documenti full-text disponibili:
[img] Documento PDF (English) - Richiede un lettore di PDF come Xpdf o Adobe Acrobat Reader
Disponibile con Licenza: Creative Commons Attribution Non-commercial ShareAlike 4.0 (CC BY-NC-SA 4.0) .
Download (35MB)


The management of a supermarket or retail store is a quite complex process that requires the coordinated execution of many different tasks (\eg, shelves management, inventory, surveillance, customer support\dots). Thanks to recent advancements of technology, many of those repetitive tasks can be completely or partially automated. One key technology requirement is the ability to understand a scene based only on information acquired by a camera, for this reason, we will focus on computer vision techniques to solve management problems inside a grocery retail store. We will address two main problems: (a) how to detect and recognize automatically products exposed on store shelves and (b) how to obtain a reliable 3D reconstruction of an environment using only information coming from a camera. We will tackle (a) both in a constrained version where the objective is to verify the compliance of observed items to a planned disposition, as well as an unconstrained one where no assumption on the observed scenes are considered. As for (b), a good solution represents one of the first crucial steps for the development and deployment of low-cost autonomous agents able to safely navigate inside the store either to carry out management jobs or to help customers (\eg, autonomous cart or shopping assistant). We believe that algorithms for depth prediction from stereo or mono camera are good candidates for the solution of this problem. The current state of the art algorithms, however, rely heavily on machine learning and can be hardly applied in the retail environment due to problems arising from the domain shift between data used to train them (usually synthetic images) and the deployment scenario (real indoor images). We will introduce techniques to adapt those algorithms to unseen environments without the need of costly ground truth data and in real time.

Tipologia del documento
Tesi di dottorato
Tonioni, Alessio
Dottorato di ricerca
Settore disciplinare
Settore concorsuale
Parole chiave
computer vision retail deep learning machine leraning object detection product recognition depth estimation stereo vision
Data di discussione
4 Aprile 2019

Altri metadati

Statistica sui download

Gestione del documento: Visualizza la tesi