High-Performance Features in Generalizable Fingerprint-Based Indoor Positioning

In the context of WiFi fingerprint-based indoor localization, the present work systematically investigates the generalization capabilities, within a k-Nearest Neighbor framework, of meta-distances learned through a genetic programming approach, considering sixteen well-known and widely used benchmark datasets. Our study reveals clear variations in meta-distances performance, emphasizing that some training datasets lend themselves to superior generalization capabilities compared to others. We identify salient features, independent from the parameter k, that appear to contribute to such a superior performance: geographically large radio-map where the average number of access points detected per fingerprint is not too high. Finally, we propose a simple approach to combine the most successful meta-distances derived from our investigation. The latter leads to an overall improvement in positioning accuracy with respect to classical distances used the literature and the previously constructed individual meta-distances, confirming the soundness of the overall experimental workflow.