Signals such as WiFi, magnetism and GPS may be detected by mobile phones and used for localization. Fingerprint-based localization, due to its deployability in complex environment, emerges as a promising approach. Because each signal has its own strengths and limitations, fusing them potentially captures their strengths while mitigating their weaknesses. Recent works on that often are highly engineered and specificically designed for two or three signals whose data have to be fully available at localization step. They can hardly be extended to embrace flexible combination of arbitrary signals with different sampling rate.

We propse SiFu, a highly accurate fingerprint-based localization framework to fuse any number and combination of heterogeneous signals. Once in operation, SiF...[ Read more ]

Signals such as WiFi, magnetism and GPS may be detected by mobile phones and used for localization. Fingerprint-based localization, due to its deployability in complex environment, emerges as a promising approach. Because each signal has its own strengths and limitations, fusing them potentially captures their strengths while mitigating their weaknesses. Recent works on that often are highly engineered and specificically designed for two or three signals whose data have to be fully available at localization step. They can hardly be extended to embrace flexible combination of arbitrary signals with different sampling rate.

We propse SiFu, a highly accurate fingerprint-based localization framework to fuse any number and combination of heterogeneous signals. Once in operation, SiFu may include new signals or exclude old ones without the need for retraining of the existing signals. To achieve this, SiFu first extracts location-dependent features from signal readings with a novel machine learning model. Based on the features, it then estimates user location with maximum likelihood estimation (MLE). SiFu is simple to implement. We conduct extensive experiments in three markedly different sites. SiFu is shown to achieve significantly better performance as compared with state-of-the-art approaches, in terms of localization error (cutting the error by 20% in our experiments).