The Evolution of Inertial Navigation and Learning-based Inertial Odometry

Inertial sensors play a pivotal role in indoor localization, which in 
turn lays the foundation for pervasive personal applications. However, 
low-cost inertial sensors, as commonly found in smartphones, are plagued 
by bias and noise, which leads to unbounded growth in error when 
accelerations are double integrated to obtain displacement. Small errors 
in state estimation propagate to make odometry virtually unusable in a 
matter of seconds. We propose to break the cycle of continuous 
integration, and instead segment inertial data into independent windows. 
The challenge becomes estimating the latent states of each window, such 
as velocity and orientation, as these are not directly observable from 
sensor data. We demonstrate how to formulate this as an optimization 
problem, and show how deep recurrent neural networks can yield highly 
accurate trajectories, outperforming state-of-the-art shallow 
techniques, on a wide range of tests and attachments. In particular, we 
demonstrate that IONet can generalize to estimate odometry for 
non-periodic motion, such as a shopping trolley or baby-stroller, an 
extremely challenging task for existing techniques.