Indoor Location Error-Detection via Crowdsourced Multi-Dimensional Mobile Data

We explore the use of multi-dimensional mobile sensing data as a means of identifying errors in one or more of those data streams. More specifically, we look at the possibility of identifying indoor locations with likely incorrect/stale Wi-Fi fingerprints, by using concurrent readings from Wi-Fi and barometer sensors from a collection of mobile devices. Our key contribution is a novel two-step process: (i) using longitudinal, crowd-sourced readings of (possibly incorrect) Wi-Fi location estimates to statistically estimate the barometer calibration offset of individual mobile devices, and (ii) then, using such offset-corrected barometer readings from devices (that are supposedly collocated) to identify likely errors in indoor localization. We evaluate this approach using data collected from 104 devices collected on the SMU campus over a period of 61 days: our results show that (i) 49% of the devices had barometer offsets that result in errors in floor-level estimation, and (iii) 46% of the Wi-Fi location estimates were potentially incorrect. By identifying specific locations with unusually high fraction of incorrect location estimates, we attempt to more accurately pinpoint the areas that need re-fingerprinting.