Abstract:
In this thesis, a practical indoor localization technique is proposed. In contrast to the state of art approaches, this practical approach does not deal with the multi path problems and shadowing effects of electromagnetic signals as well as it does not require calculating the attenuation factors for each space because it does not apply the propagation model. Instead, indoor localization, by exploiting electromagnetic scat tering properties of local area networks, is formulated as a tracking problem using a Hidden Markov model with a radio map as the observation model. Because of the non-linear relationship between radio frequency signals’ strength and location, a prob abilistic radio map is generated by using Neural Networks. Accurate estimation of the radio map is key in accurate indoor localization but this requires dense sampling of the electromagnetic field, also named as fingerprinting. To decrease the time consumption of fingerprinting process, we train the neural network using an active learning strategy based on uncertainty sampling, aided by a Gaussian process. With the radio maps generated by a deep neural network, 30% of training data can be removed and this results in an increase of 1.3% and 2.6% in median error in two different test areas. It is concluded that without trading off localization accuracy training data size can be reduced by one third.