Smartphone-based inertial and magnetic sensors can be 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 step length and heading angle are estimated to form the horizontal trajectory of the user. One of the main challenges in these methods is the unknown angular misalignment between walking direction and device orientation, which imposes great difficulty in estimating the pedestrian’s true heading.
In this work, based on accelerometer and magnetic sensors, a new framework is established to estimate the user’s heading. It comprises a novel deep network architecture, where temporal convolutions and multi-scale attention layers are trained to extract the walking direction vector in the sensors’ coordinate frame, by using acceleration signals in a rotation-invariant manner. On top of that, a unique geometric model is derived, in which gravity and geomagnetic measurements are combined with the estimated motion vector, for calculating the pedestrian’s heading angle in the north and east coordinates. The proposed model is trained and validated, in a user-dependent approach, through extensive experiments of natural walking activity with commercial smartphones.