We consider pervasive localization where a user may sample widely different signal modes (GPS,WiFi, geomagnetism, BLE, etc.) and values over time and space. Various localization algorithms have already been proposed for different signal modes. To achieve higher accuracy and pervasive localization, these signal modes may be fused. Previous works in the area are often meticulously customized for only a few (two or three) specific modes, and cannot support dynamic mode combination arisen from heterogeneous sensor sampling rates and operating conditions. Some other recent works assume a rather static environment or certain user behavior, and do not extend well to environmental changes characterized by missing signal values, signal noise, device heterogeneity, arbitrary phone carriage states...[ Read more ]

We consider pervasive localization where a user may sample widely different signal modes (GPS,WiFi, geomagnetism, BLE, etc.) and values over time and space. Various localization algorithms have already been proposed for different signal modes. To achieve higher accuracy and pervasive localization, these signal modes may be fused. Previous works in the area are often meticulously customized for only a few (two or three) specific modes, and cannot support dynamic mode combination arisen from heterogeneous sensor sampling rates and operating conditions. Some other recent works assume a rather static environment or certain user behavior, and do not extend well to environmental changes characterized by missing signal values, signal noise, device heterogeneity, arbitrary phone carriage states, etc.

We propose SiFu, a novel, highly accurate and generic multi-modal signal fusion platform supporting arbitrary addition and combination of signal modes, and robust against operational deviations from the original design point. To achieve genericity, SiFu leverages upon any existing single-modal localization algorithms as black boxes, and unifies them into a multi-modal likelihood framework. It employs Bayesian deep learning to achieve high accuracy, and data augmentation to withstand against environmental variations. Using a weighted likelihood, it fuses the modes with inertial sensor measurements by means of a particle filter. SiFu is simple to implement, and is extensible to any emerging or future signals with only incremental training. We conduct extensive experiments in three markedly different and representative sites (campus, mall and subway station), and show that SiFu achieves significantly higher accuracy as compared to other state-of-the-art approaches, cutting the localization error by more than 20% in our experiments. It is also robust against environmental variations (with 30% error reduction), even when signal values are greatly deviated from its original designed settings.