|M.Sc Student||Manos Adi|
|Subject||Statistical Learning for Pedestrian Indoor Navigation|
|Department||Department of Industrial Engineering and Management||Supervisors||ASSOCIATE PROF. Tamir Hazan|
|DR. Itzik Klein|
|Full Thesis text|
Inertial and magnetic sensors in smart devices can serve as the basis for pedestrian navigation, whenever external positioning signals are limited or unavailable. Such navigation solutions are typically accomplished by a practice known as pedestrian dead reckoning, wherein instantaneous step's length and heading angle are estimated and accumulated to form the horizontal trajectory of the user. One of the main challenges in these methods is the unknown misalignment between the user's forward axis and the device's frame, which imposes great difficulty in finding the user's heading. This research starts by deriving two methods for estimating the direction of gravity, using accelerometer and gyroscope measurements. Then, by incorporating the gravity vector, several methods to determine device's heading are proposed and investigated, using either gyroscope or magnetic sensors. In the second and central part of this work, the problem of estimating walking direction is addressed by a learning-based approach. Specifically, a novel deep network architecture is designed for extracting the motion vector in the device coordinates, using accelerometer measurements. It consists of multi-scale attention layers, combined with temporal convolutions, and involves geometrically-based regularization. Furthermore, invariance to spatial rotation of the device is embedded into the model, by a computational extension that is derived analytically. The proposed network is integrated with gravity and geomagnetic directions, in a unique geometric calculation that converts the motion vector into heading angle relative to north, without explicit usage of rotation matrices or quaternions. Extensive experiments of natural walking activity were conducted by a single pedestrian, with smartphones placed freely in various pockets. These were used for training the network-based model in a user-specific approach. Finally, the entire framework was evaluated on unseen data, comparing the deep network against a mechanistic baseline method from the literature. Using the proposed network, with merely 5500 parameters, the resulting heading errors had a median value of 9.8 degrees, which was lower by 2.5 degrees compared to the baseline method.