MRILoc: Multiresolution Indoor Localization from crowdsourced samples

Abstract

The demand of Wi-Fi fingerprinting-based indoor localization has been growing steadily for its open-source and low-cost infrastructure. Since site survey is tedious and costly, crowdsourcing is becoming a typical practice for fingerprint database construction. However, erroneous location annotation and non-uniform sample density need to be addressed in sample crowdsourcing. Furthermore, the traditional massive fingerprint databases in large and complex indoor environments require longer processing time and extensive computational complexity, limiting their efficiency and scalability. We propose a Multiresolution Indoor Localization from crowdsourced samples (MRILoc) system to address these challenges. The MRILoc consists of three major offline modules and a new online hierarchical and multiresolution localization algorithm. Firstly, we designe a Sample Reliability Measure Algorithm (SRMA) to identify reliable crowdsourced samples. Secondly, we construct weighted surfaces using reliable fingerprints and downsample fingerprints at the center of grids to resolve the issue of the non-uniform sample density. Thirdly, we construct hierarchical training databases for multiresolution localization. Our online algorithm integrates the K-Nearest Neighbors (KNN) to classify a test sample in different resolution subareas and uses XGBoost regression for the final exact localization approximation if necessary. To evaluate the performance of the proposed method, we conduct experiments on two field measurement datasets. The experiment results show a high average hitting rate and 19% localization accuracy improvement over the peer schemes.