A comparison of Appearance-based descriptors in a Visual SLAM approach
The problem of Simultaneous Localization and Mapping (SLAM) has been studied thoroughly in the past decade in the field of mobile robotics. Taking into account the information that we can find on literature, it is possible to face the SLAM problem from three different points of view: the Metric SLAM, the Topological SLAM and the hybrid Metric-Topological SLAM. When we use the metric approach, we represent the environment and we compute the robot location through geometrical information with certain accuracy. On the other hand, when we face the problem using the topological approach, the objective is to represent the environment information by means of a list of locations within a graph, maintaining connectivity relationships between them. Finally, the metric-topological approach consists of a combination of the both techniques, trying to take advantage of both methods. Nowadays, the use of computer vision is usual when we want to build a map and localize the robot in the map, because of several advantages (they are passive sensors, have a low cost and provide us with a great amount of information). When we use a vision sensor on a SLAM problem, we can approach the problem from two points of view: using the local appearance (landmarks) or using the global appearance to extract the necessary information from the scenes. The use of local appearance implies the extraction of distinctive landmarks from the images. When we use techniques based on local appearance, we typically need more computational time to build the map and locate the robot within the map. It is due to the fact that we need to extract the distinctive landmarks from each image and find each landmark extracted, in all images that compose the map. However, it presents an advantage: the possibility of including metric information to the system. Conversely, the global appearance methods need a lower time to work (they allow us to work in real time) but they do not directly include metric information in the map. The main objective of this work is to build and test an algorithm to solve the SLAM problem using the global appearance of omnidirectional visual information and the robot internal odometry. Taking into account the advantages and disadvantages of the methods previously listed, we have decided to use a hybrid metric-topological approach to solve the SLAM problem.