TOPOLOGICAL/METRICAL VISUAL SLAM BY MEANS APPEARANCE-BASE METHODS IN NON-DETERMINISTIC ENVIRONMENTS
Dr. Lorenzo Fernández Rojo

Robotics is considered as part of engineering that handles the application of infor- matics to the design and use of machines to replace people in performing different tasks. Within this area of study can be locate the present PhD thesis, more specifically in the context of mobile robotics.

If we define the term robot as an autonomous entity that has the quality of being able to move, we can say that one of the main features that should have the robot corresponds with the ability to navigate autonomously through an environment. For this purpose, three associated problems appear: Mapping, Localization y Path Planing. The term mapping may be defined as the process carried out to create a map of the environment, using data stored by robot navigation. On the other hand, the localization process can be understood as the process by which the position of the robot is estimated within a map. At last, the process followed to establish the best way forward for the robot to go from an origin point to a target point, is defined as path planning. In practice, these processes are interrelated. More specifically, creating a map while at the same time the robot is located within the same, is commonly known as the Simultaneous Localization and Mapping (SLAM) problem, and may be considered as one of the most complex tasks in the field of mobile robotics.

Regarding the type of information the robot gets of the environment, it is possible to find different proposals in the related literature (range sensors, vision sensors, etc.). Between them, it is possible to highlight the increasing use of vision sensors, due to the numerous advantages they provide. These sensors are passive sensors which provide a large amount of information without adding mechanical devices, and also has a low cost. Based on the advantages offered, for the development of this dissertation we have deci- ded to use vision sensors, more specifically, an omnidirectional vision system, providing information on the entire environment surrounding the robot (360o in the ground plane). Moreover, in contrast to most studies that can be found in the scientific literature, we have decided to use the global appearance information of the captured images, without extracting local information of these.

Based on the selected characteristics, in the present dissertation it has proposed the creation of a topological map of an environment using the global appearance information of the omnidirectional images. Topological maps are graphics models which have no metric information in Euclidean space, composed of nodes that represent a certain point in space and that, at the same time, are interconnected in a compact form. These methods need to use a descriptor that allows to recognize each captured image, and that, at the same time, decrease the computational cost (compress the captured information). With this purpose, we have decided to use the Fourier Signature as a descriptor of the images

appearance. Thus, it proposes and evaluates an algorithm to build a topological map of the environment, using only information provided by the Fourier Signature.

On the other hand, in line with the tasks that the mobile robot should be able to sol- ve, in this PhD thesis, a solution to the problem of localization is also proposed. In this way, two topological localization methods and an hybrid topological-metric localization method are proposed and evaluated. The proposed topological algorithms have been de- veloped entirely, and the metric algorithm is implemented by a modification of a Monte Carlo localization algorithm.

Once the mapping problem and the localization problem has been studied separately, arises the need to study together, because in most cases, when a robot must perform a task, this has no map or the map changes over time. With this purpose, a new SLAM algorithm is proposed in this dissertation. The SLAM algorithm we propose, is a combination of a topological algorithm with a metric algorithm.

The validity of the proposals made in this PhD thesis has been verified by means of a series of experiments. For this purpose real data, captured by mobile robots in different indoor environments, are used. Furthermore, experiments are presented in outdoor envi- ronments, also using real data. From the results presented, the validity of the proposed solutions in each of the work performed, is shown.