On Building an Indoor Radio Map from Crowdsourced Samples with Annotation Errors

Fingerprint-based indoor localization has been intensively researched in the last decade, yet the labor-intensive and time-consuming site survey for radio map construction has impeded its practical implementations. Recently, crowdsourcing has been promoted as a promising approach to exploit casually collected samples for radio map construction, however, such samples may be labelled with erroneous location information. In this paper, we study how to construct a radio map from crowdsourced samples with annotation errors. In the sample set for creating an offline fingerprint, we apply a new local density-based clustering technique to detect and remove samples being classified as outliers; Next, we propose an access point selection algorithm based on the sample signal variation to further refine each offline fingerprint structure. Field measurements and experiments are conducted and results validate the necessity of outlier removal and effectiveness of our proposed schemes. Comparing with the peer schemes, ours can achieve significant improvements of localization performance.