Estudo de métodos de mapeamento baseados em informação visual utilizando robôs móveis em ambientes confinados

Abstract

Robotics is one of the most important fields of modern technology and is increasingly present in people’s daily lives. The application of robots occurs in the most diverse areas, often related to the preservation of human health, in tasks such as inspection, extraction, maintenance, exploration, and mapping. These activities can be exhausting and dangerous for human beings, especially when the place presents risks, as is the case of confined environments. Such locations have limited entry and exit, insufficient ventilation to remove particles, and they were not even designed for continuous human occupation. In industrial and urban areas, there are several environments included in this category, such as dams, vessels, tunnels, caves, pipes, and underground galleries, which require constant inspection and maintenance to ensure correct operation. The Instituto Tecnol´ogico Vale (ITV), in partnership with the Universidade Federal de Minas Gerais (UFMG), has been developing the EspeleoRobˆo, whose main purpose is to reduce risks related to human presence in hazardous areas. One of the most important tasks of EspeleoRobˆo is the map generation of the environments, which is essential for the development of several applications in mobile robotics, such as localization, navigation, path planning, and possible inspection and maintenance procedures. Cameras stand out among the sensors used in the reconstruction of environments, being able to acquire much visual information from the surroundings. In this context, the objective of this thesis is to investigate methods of visual reconstruction able to generate dense, colored, and accurate maps in confined environments, for further purposes of inspection and structural analysis by specialists in speleology, using digital and virtual reality models. For this aim, two methods are presented: the Point Cloud Registration, combined with an external localization with an Extended Kalman Filter (EKF), and the Simultaneous Localization and Mapping (SLAM) process with Real-Time Appearance-Based Mapping (RTAB-Map), both using RGB-D cameras. The localization and mapping results obtained with the methods are analyzed and compared in simulated and real environments, in addition to their respective processing times. In the end, the results in simulation of both methods present values very close to the reference, and, for real environments, the RTAB-Map shows to be more suitable, generating more adequate reconstructions and closer to the reference.