Egocentric Vision Based Localization of Shopping Cart

Indoor camera localization from egocentric images is a challenge computer vision problem which has been strongly investigated in the last years. Localizing a camera in a 3D space can open many useful applications in different domains. In this work, we analyse this challenge to localize shopping cart in stores. Three main contributions are given with this thesis. As first, we propose a new dataset for shopping cart localization which includes both RGB and depth images together with the 3-DOF data corresponding to the cart position and orientation in the store. The dataset is also labelled with respect to 16 different classes associated to different areas of the considered retail. A second contribution is related to a benchmark study where different methods are compared for both, cart pose estimation and retail area classification. Last contribution is related to the computational analysis of the considered approaches.