Uso de descriptores holísticos para la localización y creación de mapas: una aproximación al graph-SLAM mediante apariencia visual
Throughout the last few years, the quantity of applications that use mobile robots have increased and they are present in very diverse fields. A robot must have an internal representation of the environment through which it is going to move. This representation will allow it to estimate its position...
| Autor: | |
|---|---|
| Tipo de recurso: | tesis doctoral |
| Fecha de publicación: | 2018 |
| País: | España |
| Institución: | Universidad Miguel Hernández de Elche |
| Repositorio: | REDIUMH. Depósito Digital de la UMH |
| OAI Identifier: | oai:dspace.umh.es:11000/5110 |
| Acceso en línea: | http://hdl.handle.net/11000/5110 |
| Access Level: | acceso abierto |
| Palabra clave: | Robótica Inteligencia artificial Tratamiento digital de imágenes CDU::6 - Ciencias aplicadas::62 - Ingeniería. Tecnología CDU::6 - Ciencias aplicadas::68 -Tecnología cibernética y automática |
| Sumario: | Throughout the last few years, the quantity of applications that use mobile robots have increased and they are present in very diverse fields. A robot must have an internal representation of the environment through which it is going to move. This representation will allow it to estimate its position and orientation, as well as the trajectory to follow during the operation. In this way it can carry out tasks, autonomously, within that environment. This information is essential to generate the necessary signals to feed the equipped actuators. These actuators generate the desired movement in the robot. This environment representation is done by collecting information from the surroundings through different sensors. There are many types of sensors such as lasers, encoders, GPS, sonars, visual sensors, etc... All of them provide information to the robot. The robot uses these information to generate a representation of the environment. Among all of them, it is possible to highlight the visual sensors, due to different advantages such as its reduced weight, its limited energy consumption and, above all, its multiple possibilities of configuration, which makes it a suitable sensor for an infinity types of applications. These sensors provide images that have rich information that can be exploited in different ways. Among all the possible configurations of visual sensors there is one that provides information in all directions around the robot. This configuration consists of a catadioptric system formed by a conventional camera pointing to the base of a convex mirror, which can be spherical, conical, elliptical, hyperbolic or parabolic. The entire environment is reflected in the mirror, and the camera captures this reflection, generating an omnidirectional image, which contains information about the environment with a 360 visual field around the axis of the catadioptric system. For the development of this Doctoral Thesis, the hyperbolic mirror has been chosen as a mirror of the system. This kind of mirror has different advantages that will be detailed throughout the document. The visual information obtained is very rich and this can be a problem when working with it. For this reason, the use of thechniques to extract the most relevant information is essential to do possible its management. There are different methods to perform this task in order to generate maps of the environment based on descriptors that store the relevant information. The first one is based on the extraction and description of characteristic points of the scenes, which has reached a certain maturity because it has been used in many of the known mapping and localization algorithms. However, these systems have several disadvantages due to the high computational cost to manage it and their low robustness against changes in the environment. The second method consists of working with the general information of the scenes, generating a single descriptor that collects the information of each image as a whole, without the extraction of local characteristics. It permits to extract a single holistic descriptor per image that collects its global information. This approach is more recent than the previous one and, normally, leads to simpler localization algorithms, conceptually speaking. Due to the immaturity of these methods, it is necessary to carry out exhaustive studies to prove their validity in tasks of mapping and location. In this Doctoral Thesis, localization and mapping techniques are developed using this image description approach. With this approach in mind, various algorithms, collected in each of the chapters of the Doctoral Thesis, have been proposed. First, a program to generate omnidirectional images in virtual environments is proposed to solve the problem of acquiring omnidirectional images with a real catadioptric system. This program permits to confgurate the catadioptric system by testing it using virtual images. This saves on costs and time because the acquisition of different actual databased using different system configurations is not necesary. Secondly, two 2D localization algorithms are proposed. they use a descriptor based on the Radon transform, starting from different initial hypotheses. These algorithms are compared with methods based on the extraction of characteristics and with methods based on other known global appearance descriptors. Third, a relative height estimation algorithm has been developed using the global appearance of the scenes and compared with an alternative method based on the description of the local characteristics of the scenes. Finally, a method of graph-SLAM (Creation of topological maps and localization simultaneously) is proposed. All the developed algorithms have been validated through experiments that use diverse databases formed by omnidirectional images.- |
|---|