Case Study of Deep Learning Methods for Depth Estimation in Indoor Ground Robotics

• Fábio Leandro Vizzotto CPQD
• Marcos D'Addio de Moura CPQD
• Vinicius Carbonezi de Souza CPQD https://orcid.org/0009-0006-1820-2825
• Cides Semprebom Bezerra CPQD
• Guilherme Ribeiro Sales CPQD
• Valentino Corso CPQD
• Luiz Eduardo Pita Mercês Almeida CPQD https://orcid.org/0009-0009-6823-1570
• Douglas Henrique Siqueira Abreu PUC-Campinas https://orcid.org/0009-0005-4739-5980

Abstract

Depth estimation is the computer vision task that assigns a distance between the camera and each pixel in an image. This paper focuses on monocular metric depth estimation in videos, which infers a distance in metric units using a single RGB camera. Considering its applications, robotics systems and environmental mapping arise as practical areas that can make extensive usage of these techniques. As a case study for indoor robotics, the ICL ground robot dataset obtained by video footage in graphic simulation was used for experiments. A comparison was made considering the results and requirements of data acquisition needed for different deep learning models, presenting self-supervised and supervised methods available in literature and being the first work to present a depth estimation benchmark for the chosen dataset.

Keywords: Depth estimation, Ground robotics, Deep learning, Case Study