Mapeamento de trafegabilidade baseado em fusão de dados inerciais e nuvens de pontos

• Ano de defesa: 2020
• Autor(a) principal: Felipe Gomes de Oliveira
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Minas Gerais; Brasil; ICX - DEPARTAMENTO DE CIÊNCIA DA COMPUTAÇÃO; Programa de Pós-Graduação em Ciência da Computação; UFMG
• Palavras-chave em Português: Mapeamento de Terreno; Navegação Autônoma; Aprendizado Profundo; Robótica de Campo; Computação – Teses; Mapeamento de terreno – Teses; Aprendizado profundo – Teses; Robótica – Teses
• Link de acesso: http://hdl.handle.net/1843/36848; https://orcid.org/0000-0002-5435-0933

With the growing interest in the development of autonomous vehicles for outdoor environments, it is necessary to investigate techniques that support autonomous navigation. Autonomous navigation has been widely studied by the academic community and several factors that provide a safe and efficient displacement. For autonomous navigation, they are often considered only obstacles in the environment. However, unknown and unstructured terrains may represent a crucial feature for the robot’s security or viability of the task. This work addresses the problem of mapping the difficulty level when navigating through outdoor environments from multi-sensor fusion using deep learning. In this work are considered terrains, where difficulties can be found, such as: i) different types of surfaces; ii) roughness levels disparities; and iii) highly sloping surfaces. In this way, the main objective is to create three-dimensional (3D) maps augmented with navigation costs, improving the decision making of path planning algorithms. The proposed methodology in this thesis is divided into three main steps: i) Three-dimensional mapping and localization, where is created a 3D map from point clouds provided by a laser; ii) Navigation cost estimation using inertial data, where the navigation costs are computed from inertial data provided by an IMU; and iii) 3D map augmentation with navigation cost using deep learning, where inertial and geometric data are combined through deep learning to estimate the navigation costs of unvisited regions by the ground robot. Several experiments were carried out with real robots in different environments to evaluate the quality of the proposed tasks and the complete process of navigation cost mapping. In the end, the achieved results at each proposed step are discussed.