5.3.3. Localization When Several Maps Are Available

In some applications, several maps of some different environments are initially available. If the robot has no information about the environment it is located in, first, it has to use the visual information to select the correct environment. After that, the localization can be solved in the selected environment, as presented in Section 4. Considering this, in this section, the ability to select the right environment

is studied. In order to check the goodness of the descriptors for this purpose, the two COLD maps built in Section 5.2.2 are considered. Additionally, a test dataset is created as a combination of images from the Freiburg and Saarbrüken environments. A total of 60 test images compose the test dataset (34 from Freiburg and 26 from Saarbrücken). In this experiment, only HOG and *gist* are tested again. Furthermore, since the cosine distance presented the best solutions for COLD, only this kind of distance is applied. Figure 19 shows the percentage of success in selecting the right environment for the two descriptors.

By and large, the correct environment selection is almost always done. Many cases are given in which 100% success is reached, whereas the worst cases do not present a success rate under 75%. If the environment selection is carried out with HOG, results depend substantially on the chosen *k*<sup>2</sup> value. For instance, the worst cases are presented for *k*<sup>2</sup> = 2, 4. However, for *k*<sup>2</sup> = 32 − 128, 100% success is reached. Through the use of *gist* descriptor, 100% success is given independently of the number of clusters or the *k*<sup>3</sup> value.

**Figure 19.** Percentage of success to detect the correct environment between Freiburg and Saarbrücken with FS, HOG, and *gist* used to describe the representatives of the clusters and the test images: percentage of success vs. number of clusters.
