Learning-Based Floor-Type Identification in the ZUPT-Aided Pedestrian Inertial Navigation

We present a method to identify the floor types in the pedestrian inertial navigation with the goal of adaptive estimation of navigation states under various navigation scenarios. The floor-type detection algorithm includes three major steps. First, the inertial measurement unit readouts are divided with each partition corresponding to a full human gait cycle. Second, principal component analysis is conducted on each partition to reduce dimensionality of the data. Third, the data are fed into an artificial neural network (ANN) for the floor-type identification. In this study, we identified five different floor types: walking on hard floor, walking on grass, walking on sand, walking upstairs, and walking downstairs, with an accuracy of over 99%. To verify the effects of the floor-type detection, the classification result was used to select proper parameters in the zero-velocity-update-aided pedestrian inertial navigation in the multiple model approach. Advantages of the identification approach were experimentally verified, and resulted in a reduced navigation error in the field test.