Machine Learning for Indoor Localization Without Ground-truth Locations

Wi-Fi Received Signal Strength Indicator (RSSI) has been considered as one of most important resources for indoor localization. However, the conventional RSSI fingerprint method needs recording many ground-truth locations. In this paper, we propose an unsupervised Wi-Fi RSSI learning without true location. Pseudo-labels are made by another machine learning model, which uses inertial measurement unit (IMU) as training input. From the experiments in an office building, the proposed algorithm validates to learn an unsupervised Wi-Fi RSSI positioning model by resulting in 0.8 CDF at 2 meter distance.