Improved Window Segmentation for Deep Learning Based Inertial Odometry

Abstract

The variety of sensors embedded in smartphones makes it possible to develop indoor navigation and localization systems on mobile terminals. However, these cheap sensors are plagued by bias and noise, leading to unbounded system drifts. Inspired by Expectation-Maximization algorithm, this paper proposes to combine zero-velocity detection with gated recurrent unit (GRU) neural networks, make full use of pedestrian motion characteristics, and naturally and accurately split the raw measurements into multiple weakly correlated windows step by step. The GRU is used to exploit dynamic context and predict the polar vector of each window. Several experiments were conducted to test the performance of proposed model, and IONet, a deep learning based inertial odometry model using fixed-size sliding window, was taken as a reference. The results show that the proposed model is able to generate smooth trajectories with high precision. Compared with IONet, the performance of proposed model in turning is better.