IMPROVED INDOOR POSITIONING SYSTEM PERFORMANCE USING MACHINE LEARNING

Abstract

Indoor positioning refers to the act of locating assets within an indoor space. Indoor Positioning Systems (IPS) can provide real-time location data which can improve planning, enhance productivity, and increase profits. However, their use of wireless technologies makes them vulnerable to interference and significant attenuation caused by clutter and nonideal environmental conditions in the location, diminishing signal quality and localization accuracy. Recent machine learning solutions in literature can improve accuracy, but they do not often consider practical constraints, which leads to a discrepancy between the reported accuracy during trials and the achievable accuracy during deployment. This thesis therefore develops practical machine learning solutions to improve IPS accuracy and robustness using three approaches: replacement, where a trained model acts as the positioning engine; extension, where machine learning is utilized in areas with insufficient localization setups or harsh wireless environments; and augmentation, where data-driven modules enhance the traditional localization algorithm.