Deep LSTM-Based Multimode Pedestrian Dead Reckoning System for Indoor Localization

We propose a multimode pedestrian dead reckoning (PDR) system using a recurrent neural network. We adopt a long short-term memory (LSTM) layer to extract latent features from the sensor data. We then transform the extracted latent vector using conditional input of the pedestrian's mode to make the model operate in different contexts. Finally, the step length and heading angle are obtained through a multilayer neural network with the transformed sensor latent vector as input. The simulation results show that the proposed scheme can track pedestrians in the multimode situation using a single model.