HybLoc: Hybrid Indoor Wi-Fi Localization Using Soft Clustering-Based Random Decision Forest Ensembles

Indoor localization has garnered the attention of researchers over the past two decades due to diverse and numerous applications. The existing works either provide room-level or latitude-longitude prediction instead of a hybrid solution, catering only to specific application needs. This paper proposes a new infrastructure-less, indoor localization system named HybLoc using Wi-Fi fingerprints. The system employs Gaussian Mixture Model (GMM)-based soft clustering and Random Decision Forest (RDF) ensembles for hybrid indoor localization i.e., both room-level and latitude-longitude prediction. GMM-based soft clustering allows finding natural data subsets helping cascaded classifiers better learn underlying data dynamics. The RDF ensembles enhance the capabilities of decision trees providing better generalization. A publically available Wi-Fi fingerprints data set UJIIndoorLoc (multi-floor and multi-building) has been used for experimental evaluation. The results describe the potential of HybLoc to provide the hybrid location of user viz a viz the reported literature for both levels of prediction. For room estimation, HybLoc has demonstrated mean 85% accuracy, 89% precision as compared with frequently used k Nearest Neighbors (kNN) and Artificial Neural Network (ANN)-based approaches with 56% accuracy, 60% precision and 42% accuracy, 48% precision, respectively, averaged over all buildings. We also compared HybLoc performance with baseline Random Forest providing 79% accuracy and 82% precision which clearly demonstrates the enhanced performance by HybLoc. In terms of latitude-longitude prediction, HybLoc, kNN, ANN, and baseline Random Forest had 6.29 m, 8.1 m, 180.7 m, and 10.2 m mean error over complete data set. We also present useful results on how number of samples and missing data replacement value affect the performance of the system.