SLAM Monocular en tiempo real
Cargando...
Autores
Acosta Amaya, Gustavo Alonso
Director
Tipo de contenido
Document language:
Español
Fecha
Título de la revista
ISSN de la revista
Título del volumen
Documentos PDF
Resumen
Una de las tareas fundamentales que debe poder ejecutar un robot móvil es la navegación autónoma de su entorno de trabajo, para lo que se requiere de un modelo del entorno o mapa y un método para estimar la localización en éste. Sin embargo son numerosas las situaciones en las que no se dispone, a priori, de una representación del entorno, por ejemplo en labores de búsqueda, rescate, exploración planetaria, exploración oceánica y minería subterránea. En tales circunstancias se deberán resolver de manera simultánea ambos problemas, la localización y el mapeo. En efecto, la estimación de la localización requiere de un mapa, y a su vez, para elaborar un mapa es necesario establecer una localización con relación a un modelo. La solución simultánea de estos dos problemas se conoce en robótica como SLAM (Simultaneous Localization and Mapping). Desde su formulación hace más de treinta años, la comunidad científica vinculada a la robótica ha invertido esfuerzo y energía en la solución del SLAM. En la actualidad se considera un componente fundamental de los sistemas robóticos, permitiéndoles realizar tareas más complejas y por tanto, otorgándoles mayores niveles de autonomía. Si bien el SLAM bidimensional para entornos interiores de pequeña escala se considera un problema resuelto con el que se obtienen resultados consistentes, una vez se extiende a la estimación y reconstrucción tridimensional o a entornos de grandes dimensiones, nuevos retos de investigación surgen de inmediato. Para el caso particular del SLAM tridimensional con sensores visuales, algunos de los nuevos problemas que deben ser resueltos son: mayor complejidad y costo computacional debido al gran volumen de datos a procesar, errores debidos a la baja resolución de los sensores, cambios de iluminación en el entorno, superficies con falta de textura e imágenes borrosas por movimientos ápidos de la cámara. En esta tesis se adelanta un estudio sistemático y riguroso del SLAM, desde su formulación y métodos de solución, hasta la evaluación de algunos de los algoritmos SLAM de código abierto más recientes. De manera particular se aborda el problema del SLAM monocular en tiempo real y se conducen experimentos en entornos interiores con un sistema robótico especialmente diseñado para tal fin. Las principales contribuciones de este trabajo son: - El estudio sistemático y exhaustivo del SLAM, desde su formulación hasta los métodos de solución más representativos (Capítulo 2). - La formulación de un método de evaluación de los algoritmos SLAM con base en dos métricas (MeC y MoC) que consideran la calidad de los mapas producidos (Capítulo 3). - El diseño y construcción de un sistema robótico totalmente compatible con ROS (Robot Operating System) para la validación experimental conducida en esta tesis y la investigación y desarrollo de aplicaciones (Capítulo 3). - El estudio riguroso de los algoritmos visual SLAM que hacen parte del estado del arte actual y, de manera particular, los métodos relativos a los problemas fM (Structure from Motion) y el SLAM monocular en tiempo real (Capítulos 4 y 5).
Abstract: One of the fundamental tasks that a mobile robot must be able to execute is the autonomous navigation of its working environment, for which a model of the environment or map is required and a method to estimate the location in it. However, there are numerous situations in which there is not available, a priori, a representation of the environment, for example in search and rescue, planetary exploration, ocean exploration and underground mining. In that circumstances, both problems, location and mapping must be solved simultaneously. In effect, location estimation requires a map, and in turn, to create a map it is necessary to establish a location in relation to a model. The simultaneous solution of these two problems is knows in robotics as SLAM (Simultaneous Localization and Mapping). Since its formulation more than thirty years ago, the research community linked to robotics has invested effort and energy in the solution of SLAM. Nowadays, it is considered a fundamental component of robotic systems, allowing them to perform more complex tasks and therefore, giving them greater levels of autonomy. Although the two-dimensional SLAM for small-scale indoor environments is considered a solved problem, with consistent results, when it extends to three dimensional estimation and reconstruction or to large-scale environments, new research challenges emerge immediately. Particularly, for the three-dimensional SLAM with only visual sensors, some of the new problems that must be solved are: greater complexity and computational cost due to the large volume of data to be processed, errors due to low resolution of the sensors, lighting changes in the environment, surfaces with lack of texture and blurry images due to rapid movements of the camera. In this thesis a systematic and rigorous study of the SLAM is carried out, from its formulation and solution methods, until the evaluation of some of the most recent open source SLAM algorithms. In particular, the monocular SLAM problem in real time is addressed and experiments are conducted in indoor environments with a robotic system specially designed for this purpose. The main contributions of this work are: - A systematic and exhaustive study of SLAM, from its formulation to the most representative solution methods, is conducted (Chapter 2). - The formulation of a method for evaluating SLAM algorithms based on two metrics (MeC y MoC) that consider the quality of the maps produced (Chapter 3). - The design and construction of a robotic system with full compatibility with ROS (Robot Operating System) for the experimental validation conducted in this thesis and research and development of applications (Chapter 3). - The rigorous study of visual SLAM algorithms that are part of the current state of the art and, in particular, the methods related to SfM (Structure from Motion) and real-time monocular SLAM (Chapters 4 and 5).
Abstract: One of the fundamental tasks that a mobile robot must be able to execute is the autonomous navigation of its working environment, for which a model of the environment or map is required and a method to estimate the location in it. However, there are numerous situations in which there is not available, a priori, a representation of the environment, for example in search and rescue, planetary exploration, ocean exploration and underground mining. In that circumstances, both problems, location and mapping must be solved simultaneously. In effect, location estimation requires a map, and in turn, to create a map it is necessary to establish a location in relation to a model. The simultaneous solution of these two problems is knows in robotics as SLAM (Simultaneous Localization and Mapping). Since its formulation more than thirty years ago, the research community linked to robotics has invested effort and energy in the solution of SLAM. Nowadays, it is considered a fundamental component of robotic systems, allowing them to perform more complex tasks and therefore, giving them greater levels of autonomy. Although the two-dimensional SLAM for small-scale indoor environments is considered a solved problem, with consistent results, when it extends to three dimensional estimation and reconstruction or to large-scale environments, new research challenges emerge immediately. Particularly, for the three-dimensional SLAM with only visual sensors, some of the new problems that must be solved are: greater complexity and computational cost due to the large volume of data to be processed, errors due to low resolution of the sensors, lighting changes in the environment, surfaces with lack of texture and blurry images due to rapid movements of the camera. In this thesis a systematic and rigorous study of the SLAM is carried out, from its formulation and solution methods, until the evaluation of some of the most recent open source SLAM algorithms. In particular, the monocular SLAM problem in real time is addressed and experiments are conducted in indoor environments with a robotic system specially designed for this purpose. The main contributions of this work are: - A systematic and exhaustive study of SLAM, from its formulation to the most representative solution methods, is conducted (Chapter 2). - The formulation of a method for evaluating SLAM algorithms based on two metrics (MeC y MoC) that consider the quality of the maps produced (Chapter 3). - The design and construction of a robotic system with full compatibility with ROS (Robot Operating System) for the experimental validation conducted in this thesis and research and development of applications (Chapter 3). - The rigorous study of visual SLAM algorithms that are part of the current state of the art and, in particular, the methods related to SfM (Structure from Motion) and real-time monocular SLAM (Chapters 4 and 5).