.png){kind=icon}
.png){kind=icon}
Return
Revisiting Deep Inertial Odometry: Two Complementary Methods for Drift Reduction
.png){kind=icon}
PRE
AI
DOI：10.1109/JSEN.2025.3634088.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
In this work, we address the problem of drift mitigation in the estimation of trajectories from inertial data using deep neural networks. We propose two methods to reduce the error. The first method applies the empirical wavelet transform (EWT) algorithm to the trajectories estimated by various deep-learning-based approaches (6DOF Odometry, IONet, RoNIN, and VRIO). The second method shifts the reference frame from local to global, eliminating the need for orientation estimation. We leverage multiple publicly available datasets (OxIOD, EUROC, RIDI, RoNIN, and VRD), all of which provide inertial readings from an inertial measurement unit (IMU) alongside corresponding ground-truth pose measurements. Our experiments demonstrate that EWT can successfully reduce position estimation error by removing low-frequency drift components (i.e., mode zero) from the estimated trajectory using the neural network. In the OxIOD and VRD datasets, subtracting mode zero aligns the estimated trajectory with the ground truth and reduces the mean absolute error (MAE). However, in other datasets, improvement is mainly observed along the gravity axis, while performance might degrade on other axes. For evaluating the second method, we modify the 6DOF Odometry and IONet frameworks to operate in a global reference frame, removing the need to estimate the orientation, which simplifies the learning target and the loss function design, and yields better accuracy in several datasets.
Keywords:
Deep learning
drift reduction
empirical wavelet transform (EWT)
inertial measurement unit (IMU)
inertial odometry
trajectory estimation
Journal
IF:
4.5
Papers:
2.1W
Citations:
7.3W