Transferring Localization Models over Time

Abstract

Learning-based localization methods typically consist of an offline phase to collect the wireless signal data to build a statistical model, and an online phase to apply the model on new data. Many of these methods treat the training data as if their distributions are fixed across time. However, due to complex environmental changes such as temperature changes and multi-path fading effect, the signals can significantly vary from time to time, causing the localization accuracy to drop. We address this problem by introducing a novel semi-supervised Hidden Markov Model (HMM) to transfer the learned model from one time period to another. This adaptive model is referred to as transferred HMM (TrHMM), in which we aim to transfer as much knowledge from the old model as possible to reduce the calibration effort for the current time period. Our contribution is that we can successfully transfer out-of-date model to fit a current model through learning, even though the training data have very different distributions. Experimental results show that the TrHMM method can greatly improve the localization accuracy while saving a great amount of the calibration effort.