Next: Landmark-based localization Up: Localization Previous: Localization
These algorithms extract geometric features from the sensor readings and try to match them with a map of the environment, in order to correct possible odometric errors. This approach is closely related to grid-based mapping (described below), since these geometric features are the information pieces that grid-based mapping approaches store on the map.
The position of the robot is incrementally computed using odometry and information from sensors, by matching this information with the map already built. The sensor information used for matching can be single sonar scans, which are matched with the obstacles on the map, such in Moravec and Elfe's approach [23,52]. Other approaches, such as Chatila and Laumond's [18] extract geometric features (line segments and polyhedral objects) from the sensor readings and match them to a geometric map of the environment.
One problem with this approach is that it requires accurate odometry to disambiguate among positions that look similar. Probabilistic approaches (Smith et al [61], Fox et al [27], Castellanos and Tardós [16]) try to solve this ambiguity problem, and they are the most frequently used in the field of robot mapping. The basic idea of these algorithms is to employ probabilistic models of the robot and the environment to cope with the uncertainty of robot motion and sensor reading. In order to localize the robot, they use consecutive sensor readings to estimate a distribution over the space of all locations in the environment. The more readings the robot gets, the more precisely its location can be computed.
In our case we do not have to deal with this ambiguity, since we have developed a Vision system robust enough to correctly identify the landmarks. Thus, there is no uncertainty about the presence of a landmark. However, there is some imprecision about its location, which we deal with using fuzzy techniques.
The model matching approach, however, is computationally very expensive, since the process of matching the current sensor readings with the map requires many computations.
© 2003 Dídac Busquets