Evaluation of Aligning Methods for Landmark-Based Maps in Visual SLAM

Abstract

Building maps is one of the most fundamental tasks for an autonomous robot. This robot should be able to construct a map of the environment and, at the same time, localize itself in it. This problem, known as Simultaneous Localization and Mapping (SLAM), has received great interest in the last years (Leonard & Durrant-Whyte, 1991). In our particular case, the robots build their maps using the FastSLAM algorithm (Montemerlo et al., 2002). The main idea of the FastSLAM algorithm is to separate the two fundamental aspects of the SLAM problem: the estimate of the robot’s pose and the estimate of the map. This algorithm uses a particle set that represents the uncertainty of the robot’s pose (localization problem) meanwhile each particle has its own associated map (several individual estimates of the landmarks conditioned to the robot’s path). The solution to the SLAM problem is performed by means of a sampling and particle generation process, in which the particles whose current observations do not fit with their associated map are eliminated. The FastSLAM algorithm has proved to be robust to false data association and it is able to represent models of non-linear movements in a reliable way (Howard, 2006). In relation to the sensors used to build the maps, many authors use range sensors such as sonar (Kwak et al., 2008; Wijk & Christensen, 2000) or laser (Thrun, 2004; Triebel & Burgard, 2005). Nevertheless, there is an increasing interest on using cameras as sensors. This approach is denoted as visual SLAM (Valls Miro et al., 2006). These devices obtain a higher amount of information from the environment and they are less expensive than laser as well. Furthermore, 3D information can be obtained when using stereo cameras. These are the sensors used in this work. Most visual SLAM approaches are landmark-based. These landmarks consist of a set of distinctive points which are referred to a global reference system. The main advantage of landmark-based maps is the compactness of their representation. By contrast, this kind of maps requires the existence of structures or objects that are distinguishable enough. The map building problem can be solved by a single robot (Moutalier & Chatila, 1989), but it will be more efficiently managed if there is a team of robots, which collaborates in the construction of the map (Howard, 2006). In this case, the space can be divided so that the distances traversed are shorter and thus the odometry errors will be smaller. In this work, 7