Beyond the RSSI value in BLE-based passive indoor localization: let data speak

In this paper we present the results obtained from a large experimental environment that makes use of Bluetooth Low Energy (BLE) as the core technology for a location estimation system. BLE is a common technology for this kind of geopositioning systems, but most of the existing proposals are based on the RSS (Received Signal Strength) value obtained by mobile smart devices from static emitters such as iBeacons or other similar tags. This is not our case, since we adopt a passive approach where monitors obtain advertising frames emitted by mobile BLE beacons with no computing capabilities. In our particular scenario, based on a commercial application of our system, we perform fast but exhaustive training procedures to produce an initial dataset that is then analyzed paying attention to important design parameters. A thorough analysis of the data by means of different data visualization techniques reveals valuable information about the behavior of the emitters, signal characterization, radio coverage and, mainly, possible features that can be employed lately by machine learning methods in order to provide accurate location estimations. This is useful to define a quick and continuous training life-cycle which enables the detection of inconsistent data or failures. Our analysis also suggests that a representation of the observations using alternative features provides similar or even better results than the RSS values in terms of efficiency and support for heterogeneous devices.