6.2. Localization 876.2.3 Real World MeasurementsFor the evaluation of the algorithms under reals world conditions, the evaluationcorpus, described in table 6.5, that is composed of the defined paths in table 6.4 areused. The scene and the methods for obtaining the corpus were carefully describedin section 6.2.1.The first experiment is conducted on the evaluation samples from the Iconia device.160 samples were taken from the evaluation corpus. This was done by selecting10 samples for each defined path. This relates to an evaluated distance of around4000 m. Furthermore, an optimized radio propagation model is used, trained onlyon the measurements for the Iconia device.Figure 6.3 Localization errors for the different algorithms on the real world mea-surements from the Iconia device.As can be seen in figure 6.3, the ranking of the errors for the different approaches arecomparable to the results of the synthetics evaluation(figure 6.2) although the onlinevariants of the HMM show higher errors than their PF counterparts. The LMSEapproach is outperformed by the PF, which is less accurate in the offline variantof the error than the HMM. The reduction of the error rates of the PF and HMMwith respect to the LMSE approach are less satisfying than under the syntheticconditions. Under synthetic conditions with a noise of 14 dBm as in table 6.6, theLMSE is outperformed by a factor nearly of 2 by the HMM and PF algorithm. Underthe real world conditions of this setting, the margin is reduced to around 20%. Fromthe details of the results for the different paths separated in table 6.7 can be seen,that some of the defined paths are predicted quite inaccurately. This observation canalso be derived from the elevated standard deviation of 45 cm and upwards, whereasthe compared synthetic result show an upper limit of 45 cm(without the LMSE).Especially the stairs-upward path shows an unsatisfying offline error of around 4 mfor the HMM/PF algorithm. Whereas under synthetic condition with 14 dBm noise,it is only around 2. 5 m. This can probably be explained by the relatively low APcoverage of around 4 APs as reported in table 6.4. Another outlier is given by theog-eg path. Removing both paths from the corpus reduces the average error ratesover all algorithms by nearly 40 cm.Therefore, it can be concluded, that the absolute error rates reported in this evalu-ation are very scene specific. As such, they cannot be used for a meaningful com-parison with the systems found in literature.