SpaceYNet: a regression pose and depth-scene simultaneously

• Ano de defesa: 2020
• Autor(a) principal: Aragão, Dunfrey Pires
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Universidade Federal da Paraíba; Brasil; Informática; Programa de Pós-Graduação em Informática; UFPB
• Palavras-chave em Português: Conjunto de dados; Depth-scene; Pose; Rede neural; Regressão; Robô; Dataset; Neural network; Regression; Robot; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO
• Link de acesso: https://repositorio.ufpb.br/jspui/handle/123456789/18483

One of the fundamental dilemmas of mobile robotics is the use of sensory information to locate an agent in geographic space and avoid acting upon unknown environments that may lead to unexpected effects. In this work, we developed a global relocation system aiming to predict the robot position and avoid unforeseen actions from a monocular image, which we named SpaceYNet. Thus, Inception layers were incorporated into symmetric layers of down-sampling and up-sampling to solve depth-scene and 6-DoF estimation simultaneously. In addition, we implemented a recurrent solution to learn the contextual features of the environment. We also compared SpaceYNet to PoseNet and ContextualNet, being states of the art in robot pose regression, using CNN and LSTM in order to evaluate it. The comparison comprised one public dataset and one created in a large indoor environment. SpaceYNet showed higher accuracy in global percentages when compared to PoseNet in regressing 6- DoF, being better in 58.42% cases with the public dataset in global percentages when compared to the PoseNet, and 56.06% cases working with the newer dataset. When compared to SpaceYNet v.2 to the ContextualNet network, it also presented advantage of accuracy and comparing SpaceYNet v.1 and SpaceYNet v.2, and the difference is 67,5% of higher accuracy.