Unsupervised Learning for Crowdsourced Indoor Localization in Wireless Networks

Wireless Local Area Network (WLAN) location fingerprinting has become a prevalent approach to indoor localization. However, its widespread adoption has been hindered by the need for manual efforts to collect location-labeled fingerprints for the calibration of a localization model. Several semi-supervised learning methods have been applied to reduce such manual efforts by exploiting unlabeled fingerprints, but they still require some amount of labeled fingerprints for initializing the learning process. In this research, in order to obviate the need for location labels or references, we propose a novel unsupervised learning method that calibrates a localization model using unlabeled fingerprints based on a hybrid global-local optimization scheme. The method determines the optimal placement of fingerprint sequences on an indoor map, under the constraint imposed by the inner structure shown on the map such as walls and partitions. An efficient interaction between a global and a local optimization in the hybrid scheme drastically reduces the complexity of the learning task. Experiments carried out in a single- and a multi-story building revealed that the proposed method could successfully build a precise localization model without any location reference or explicit efforts to collect labeled samples.