K-Means Fingerprint Clustering for Low-Complexity Floor Estimation in Indoor Mobile Localization

Indoor localization in multi-floor buildings is an important research problem. Finding the correct floor, in a fast and efficient manner, in a shopping mall or an unknown university building can save the users' search time and can enable a myriad of Location Based Services in the future. One of the most widely spread techniques for floor estimation in multi-floor buildings is the fingerprinting-based localization using Received Signal Strength (RSS) measurements coming from indoor networks, such as WLAN and BLE. The clear advantage of RSS-based floor estimation is its ease of implementation on a multitude of mobile devices at the Application Programming Interface (API) level, because RSS values are directly accessible through API interface. However, the downside of a fingerprinting approach, especially for large-scale floor estimation and positioning solutions, is their need to store and transmit a huge amount of fingerprinting data. The problem becomes more severe when the localization is intended to be done on mobile devices which have limited memory, power, and computational resources. An alternative floor estimation method, which has lower complexity and is faster than the fingerprinting is the Weighted Centroid Localization (WCL) method. The trade-off is however paid in terms of a lower accuracy than the one obtained with traditional fingerprinting with Nearest Neighbour (NN) estimates. In this paper a novel K-means-based method for floor estimation via fingerprint clustering of WiFi and various other positioning sensor outputs is introduced. Our method achieves a floor estimation accuracy close to the one with NN fingerprinting, while significantly improves the complexity and the speed of the floor detection algorithm. The decrease in the database size is achieved through storing and transmitting only the cluster heads (CH's) and their corresponding floor labels.