Analysing the Performance of Fingerprinting-Based Indoor Positioning: The Non-trivial Case of Testing Data Selection

Abstract

Indoor positioning methods make it possible to estimate the location of a mobile object in a building. Many of these methods rely on fingerprinting approach. First, signal strength data is collected in a number of reference indoor locations. Frequently, the vectors of the strength of the signals emitted by WiFi access points acquired in this way are used to train machine learning models, including instance-based models.

In this study, we address the problem of signal strength data acquisition to verify whether different strategies of selecting signal strength data for model testing are equivalent. In the analysed case, the content of a testing data set can be created in a variety of ways. First of all, leave-one-out approach can be adopted. Alternatively, data from randomly selected points or same grid points can be used to estimate method accuracy. We show which of these and other approaches yield different accuracy estimates and in which cases these differences are statistically significant. Our study extends previous studies on analysing the performance of indoor positioning systems. At the same time, it illustrates an interesting problem of testing data acquisition and balancing the conflicting needs of collecting testing data in similar, yet different conditions compared to how training data was acquired.