Navigation based on inertial sensor data using deep learning techniques

Estimating the location of pedestrians continuously and in real-time indoors and outdoors is an important problem in medical rehabilitation, occupational health and safety as well as retail applications. When a clear view of satellites is available outdoors, the Global Positioning System (GPS) can provide accurate positions to smart phones and smart watches. GPS signals, however, are not always available especially in indoor or dense urban environments due to multi-path reflection or signal blockage by buildings. Utilizing the existing wireless infrastructures like cell-tower or wireless local area networks (WLANs) is another possibility for indoor pedestrian navigation by triangulation methods. Nevertheless, these solutions also suffer from multipath loss and similar signal problems. Strapdown inertial navigation with zero-velocity update (ZUPT) or pedestrian dead reckoning (PDR) methods provide a navigation solution based on on- board inertial measurement units (IMU) without depending on any external infrastructure. These dead reckoning methods, however, have unbounded consumer-grade IMU positioning errors due to sensor error accumulation by integration. These errors originate from finite and bounded random drift, axis misalignment, scale factor, and thermomechanical white noise. An appropriate way for continuous pedestrian positioning is to use an accurate IMU based method as a sub-system of an integrated navigation unit. The IMU-based positioning is a functional solution even in the case of a temporary signal blockage of the position-fixing subsystem such as the GPS or WLAN. Recently, it has been shown that deep learning (DL) based dead reckoning methods outperform the classical ZUPT and PDR methods in terms of the positioning accuracy. In this thesis, the state-of-the-art deep inertial odometry methods have been refined, made more accurate, smaller in memory size and latency. Recent DL-based dead reckoning methods show that deep recurrent neural networks can yield highly accurate trajectories compared with other shallow techniques without resorting to visual odometry. By refining the DL model architecture, we present a compact and robust deep inertial odometry methodology. While the root-mean-squared error (RMSE) for the estimated position decreases by 26% in our model, the number of trainable parameters and the latency of the artificial neural network (ANN) are decreased by 64% and 50%, respectively. Thus, the ANN method has become more feasible to implement on mobile devices and embedded systems. Furthermore, the proposed DL architecture is extended on drone positioning problem using IMUs. Using three different drone positioning datasets, DL architectures have been trained and tested. While drone-positioning using only IMU data is feasible, this problem presents additional challenges due to complex motion dynamics. Thus, this model can help improve positioning accuracy in applications involving mobile devices with indoor uses such as search and rescue, sports performance measurements, drone localization.