Supervisor/s Alberto Ortiz Rodriguez
An autonomous robot needs to perform some tasks in order to be independent. Some of these tasks require that the agent is able to locate itself in its environment. This process is known in mobile robotics as localization and it can be achieved by different ways according to the characteristics of the scenario where the robot is working.
One of this ways is that the robot use a world representation called map. There exists several methods to represent the agent environment. A widely used option is topological maps, where the world is modeled as an abstract graph with nodes which represent locations, and links between them that indicate possible actions to take between two concrete nodes.
The process of construction of these maps is called mapping. A number of sensors can be used to this end in order to obtain information from the environment. When a camera is the selected option, visual information needs to be described and managed. In this case, the quality of the mapping and localization processes depends on how images are described.
Within this context, this work introduces, on the one hand, an appearance-based method to create topological maps from a sequence of images; it also deﬁnes several measures that permit assessing the performance of dierent visual descriptors for mapping and localization tasks. On the other hand, this work also comprises a comparison of different view descriptors, involving real cases and different types of scenarios, and using the aforementioned measures as the ﬁgures of merit.
As will be seen, the developed framework and the measures exposed in this report can be easily extended and used to test more image descriptors in different environments. This work also shows a ﬁrst mapping and localization approach in underwater scenarios, not much explored yet.