Retail Store Visual Structure-from-Motion

Valter André Ribeiro dos Santos Piedade

Key information

Authors:

Valter André Ribeiro dos Santos Piedade (Valter Piedade)

Supervisors:

Pedro Daniel dos Santos Miraldo (Pedro Daniel dos Santos Miraldo)

Published in

10/28/2021

Abstract

3D mapping technologies have received increasing attention over the last few years. Among these technologies are Simultaneous Localization and Mapping (SLAM) and Structure-from-Motion (SfM) systems. One of the many applications of these systems is found in retail stores where robots are used to, e.g., autonomously navigate through the store to stock the shelves or detect missing products. In this thesis a SfM system to work in a retail store was implemented from scratch in C++, using the structure of the well-known ORB-SLAM2, proposed in [35], as a baseline for the developed method. The system receives images from a moving stereo camera and uses them to estimate visual odometry and to build a map of the traveled region. The pipeline is modular so that various feature detectors and several methods of visual odometry estimation can be used. Two hybrid methods for visual odometry estimation are further proposed in this thesis. One of them aims at exploiting the typical geometric structure of retail stores. Results of the developed system were obtained using acquired datasets from a retail store and the KITTI dataset, and show some ablation studies for each module of the pipeline.