Enhanced feature space clustering via spectral parameter weighting for fingerprinting based indoor localization

Fingerprinting based techniques have become a popular solution for indoor localization applications due to their robust performance compared to other approaches even in environments with non-line of sight and multipath conditions. Most of these techniques use Euclidean distance based algorithms such as k-nearest neighbors in the matching phase. This paper introduces a selective parameter weighting technique to enhance clustering conformity. The parameter weighting increases inter-cluster distance while reducing the intra-cluster distance, which in turn improves accuracy. This paper also extends this idea, to generate aggregated clusters that group existing clusters into umbrella clusters. This enables two phases of clustering. First an unknown location is matched to a region containing several grid points, after which the exact location is found within the region. Once the feature space is constructed as mentioned above, principal component analysis is used to reduce it to a set of uncorrelated radio maps which capture all the unique features inherent to each location. This work makes use of audible sound for the construction of radio maps and fingerprints. However, the concepts introduced here may easily be adapted for other types of fingerprinting such as Wi-Fi based fingerprinting etc.