Unsupervised Crowd-Assisted Learning Enabling Location-Aware Facilities

The accelerated evolution of Internet of Things (IoT) architectures and their incorporation in vehicles, buildings, or cities provide the ideal environment for the development and optimization of smart services. Under this light, positioning services that harvest location fingerprinting based on received signal strength indications (RSSIs) are widely popular due to the massive data generation that IoT settings provide. However, the labor-intensive and repetitive task of the radio map construction through offline RSSI fingerprint collection prevents such services from becoming standard equipment for future smart facilities. In this paper, we present a location-aware infrastructure that combines a broad sensing layer, edge computing, and centralized cloud federation support. Our setting gives rise to a sensing mechanism that enables in-facility crowdsourcing able to aid fingerprinting localization services. To that end, instead of extensive offline measurements, we use the facility occupants to gather unlabeled RSSI samples. To support the localization functionality, we develop a probabilistic cell-based model that is constructed by an unsupervised learning algorithm. Our black-box approach maintains the positioning accuracy regardless of changes in the underlying hardware or indoor environment. To evaluate our approach, we have deployed a multistorey facility testbed and performed an extensive real-subject trial to gather the unlabeled fingerprint dataset. The proposed unsupervised method yields average location classification accuracy of 0.8 that can rise up to 0.9 when a semi-supervised approach is considered. We also provide insights into the performance of the proposed infrastructure regarding mobility tracking, and under varying deployment scenarios.