Toward extrapolation of WiFi fingerprinting performance across environments

Out of the plethora of approaches for indoor localization, WiFi-based fingerprinting offers attractive trade-off between deployment overheads and accuracy. This has motivated intense research interest resulting in many proposed algorithms which are typically evaluated only in a single or small number of discrete environments. When the end-user's environment is not part of the evaluated set, it remains unclear if and to what extent the reported performance results can be extrapolated to this new environment. In this paper, we aim at establishing a relationship between the similarities among a set of different deployment environments and parameterizations of fingerprinting algorithms on one side, and the performance of these algorithms on the other. We hypothesize about the factors that can be used to capture the degree of similarity among environments and parameterizations of the algorithms, and proceed to systematically analyze the performance of two fingerprinting algorithms across four environments with different levels of similarity. The results show that the localization error distributions have small statistical difference across environments and parameterizations that are considered similar according to our hypothesis. As the level of similarity is decreased, we demonstrate that the relative performance of the algorithms can still be preserved across environments. For dissimilar environments, the localization errors demonstrate larger statistical differences.