Frequentist inference for WiFi fingerprinting 3D indoor positioning

Weighted k-Nearest Neighbors (WkNN) algorithms based on WiFi fingerprinting are a popular choice for 3D indoor position estimation. Performance of these schemes strongly depends however on the number of k Reference Points (RPs) used for the estimation. In this work a novel WiFi fingerprinting WkNN algorithm is proposed, that aims at improving position accuracy and robustness to variations of the value of k. The proposed algorithm relies on frequentist theory of inference combined with a measure of similarity given by the Pearson´s correlation R statistical index. The algorithm uses the p-value probabilities as defined in frequentist inference to determine the relevance of each RP. The algorithm is compared with preexisting WkNN algorithms as well as with a WkNN algorithm relying on the R index, also defined in this work. Experimental results show that the proposed algorithm leads to higher positioning accuracy and higher robustness to sub-optimal selection of the value k.