An ensemble of condition based classifiers for indoor localization

Radio frequency fingerprinting, based on Wi-Fi signals is a popular approach for indoor localization. Recently a few works have explored applicability of machine learning techniques to this problem. However, the challenging task of accurately finding the position depends on prior efforts of fingerprinting. Another challenge is that, distance sensitivity of signal strength depends on proximity to the access point. Heterogeneity of devices adds new dimension to the challenge. Existing solutions mostly aim at fingerprinting under the set of conditions considered so that training and test data can be taken under similar experimental setup. In this paper, an ensemble of condition based classifiers are designed for indoor localization that handles device heterogeneity, temporal heterogeneity and context heterogeneity (door and window open/close, presence/absence of other users in vicinity). We have created an indoor localization data set where data is collected in above mentioned dimensions. When the training and test data sets are taken under similar environmental conditions with same devices, the average classification accuracy ranges between 72.2% to 92.6%. But for a validation set which contains examples for every conditions, the classifiers achieve an average maximum accuracy of 74.8% when individual training set corresponding to specific conditions are used. To avoid these conditional dependencies we have used an ensemble of condition specific K-nearest neighbour classifiers which enables us to predict the location with 96% accuracy.