DO IONet: 9-Axis IMU-Based 6-DOF Odometry Framework Using Neural Network for Direct Orientation Estimation

With the development of mobile devices, low-cost inertial measurement unit (IMU)-based research on inertial odometry for indoor localization is being actively conducted. Inertial odometry estimates the amount of change in position and orientation relative to the initial value based on the data measured by the IMU and creates a trajectory via a multi-integration process. However, existing approaches primarily focus on estimating the position in a two-dimensional (2D) plane. Additionally, drift errors occur because the position is estimated by integrating the position change and the orientation change. Herein, we propose a novel six-degree of freedom inertial odometry framework that directly estimates the orientation to minimize drift errors. The proposed framework is a direct-orientation inertial odometry network (DO IONet) that outputs the velocity and orientation of the IMU by using linear acceleration, gyro data, gravitational acceleration, and geomagnetic data as inputs to structurally eliminate drift errors that occur during orientation calculations. DO IONet is composed of a convolutional neural network-based encoder to extract features from inertial data and decoder to extract sequential features. Structurally, it does not require initial values and has no cumulative error despite estimating orientation over tens of seconds.