Indoor topological slam using frontal computer vision
- Autores: Jaime Boal Martín Larrauri
- Directores de la Tesis: Alvaro Sánchez Miralles (dir. tes.)
- Lectura: En la Universidad Pontificia Comillas ( España ) en 2014
- Idioma: español
- Tribunal Calificador de la Tesis: José María Armingol Moreno (presid.), José Villar Collado (secret.), Ángela Ribeiro Seijas (voc.), Paulo José Monteiro Peixoto (voc.), Miguel Angel Sanz Bobi (voc.)
- Materias:
- Enlaces
- Tesis en acceso abierto en: TESEO
- Resumen
The last few years have seen a great leap forward towards autonomous mobile robots, and it is just a matter of time that they become a regular part of our lives. However, some fundamental problems need to be addressed before a robot can be assigned to any particular high-level application. One of these challenges is to provide the robot with the ability of localizing itself in a previously unvisited environment without having to supply it with a map of the area in advance. Therefore, when the robot is moved to a new area, it should incrementally build a map on its own and determine its position within it. This is known as the simultaneous localization and mapping problem (SLAM).
The map can be a very precise metric model or, alternatively, follow a topological approach that resembles human beings¿ more intuitive representation of space. We have an abstract notion of distance but are still able to determine where we are using vision to identify distinct places and the transitions among them. Hence, if we do not need to answer the question ¿Where am I?¿ with precision of millimeters and degrees, why should a robot? Moreover, in those applications in which a robot and human beings need to interact, the map should ideally be a common communication framework, so the more similar the map is to the way we structure spatial information, the easier interaction would be.
This thesis proposes a relatively computationally inexpensive solution to the SLAM problem inspired by human behavior. A forward-facing camera is the only sensor employed to make the system easily portable to a wide range of robotic platforms. By means of computer vision, the robot extracts a complementary collection of cues (vertical edges, color information, and keypoints) that focus both on the general characteristics of the scene and on the details, and employs a novel matching procedure that builds on concepts borrowed from the natural language processing field.
These features are then used to automatically identify qualitatively different locations that are susceptible of being considered a place in the map using an online topological segmentation algorithm based on the algebraic connectivity of graphs. Every time the robot arrives at a place, a Bayesian formulation is employed to decide if the robot is in a new or an already known location in order to update the topological map accordingly. As keeping track of all possibilities over time is computationally intractable, a particle filter is used to take only the most probable topologies into account.
Keywords: Computer vision ¿ Feature detection and matching ¿ Mobile robots ¿ SLAM ¿ Topological modeling of the environment
