5.3.2. Localization in the Freiburg Environment

As in the previous case (clustering task), with the aim of corroborating the results obtained in Quorum V, an evaluation of the localization task is carried out in the COLD environments. These environments present trajectory maps instead of grid maps. The two COLD environments present a similar configuration and also similar results. This way, only the results obtained in one of them are shown. Freiburg is chosen because this environment presents more rooms and also is more challenging due to the fact that the building presents many glass walls. Moreover, as Figure 16 shows, since the FS descriptor has presented the worst results, this descriptor is discarded in subsequent localization experiments. Furthermore, the Euclidean distance results are omitted in this section because it presented the worst outcomes. Figure 18 shows the average localization error (cm) obtained when HOG (first row) and *gist* (second row) are used respectively as the descriptor. The case of no compaction is also considered (*nc* = 519).

**Figure 18.** Results of the localization process with HOG and *gist* used to describe the representatives of the clusters and the test images: average localization error (cm) vs. number of clusters. Freiburg environment.

In this case, some differences are noticed between the results collected in the Quorum V environment and the results in the Freiburg environment. When the number of clusters is low (*nc* = [15, 25, 40]), the localization task presents a lower average localization error with *gist*. If this number is higher than 40, the localization error is very similar for HOG and *gist*. Comparing the results obtained with the two evaluated types of distances, no remarkable differences are found. Nevertheless, a slight improvement can be noticed when the cosine distance is used. For instance, the average error value when *nc* = 40 in HOG is lower with the cosine than with correlation.

Additionally, the value of *k*<sup>2</sup> in HOG is very important. The average error varies significantly according to it. Therefore, in order to solve the localization in an environment whose properties are similar to Freiburg or Saarbrücken (information along a trajectory), the optimal values are reached through the use of HOG descriptor with *k*<sup>2</sup> = [16, 32] and cosine distance.
