IONet: Learning to Cure the Curse of Drift in Inertial Odometry

Changhao Chen, Xiaoxuan Lu, Andrew Markham, Niki Trigoni

Abstract

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.

Introduction: The Curse of Drift

Inertial sensors (IMUs) in smartphones are essential for indoor navigation, but they suffer from a fundamental flaw known as the "curse of drift." Traditional Strapdown Inertial Navigation Systems (SINS) calculate displacement by double-integrating acceleration. Because low-cost sensors are inherently noisy, small errors grow cubically over time, rendering trajectories unusable within seconds.

To mitigate this, researchers often employ Pedestrian Dead Reckoning (PDR), which detects individual steps. While PDR limits drift to linear growth, it relies heavily on the assumption of periodic human walking. It fails significantly when motion is non-periodic—such as when a phone is placed on a stroller, a shopping cart, or a robot. IONet aims to provide a "pure" inertial odometry that is accurate, eliminates the need for hand-engineered step detectors, and generalizes to any smooth motion.

Methods: Learning Motion Physics

The core innovation of IONet is breaking the cycle of continuous integration. Instead of a single long chain of calculations, raw IMU data is segmented into independent, overlapping windows (typically 200 frames, or roughly 2 seconds).

Deep Neural Network Architecture

IONet utilizes a 2-layer Bidirectional LSTM (Bi-LSTM) with 96 hidden nodes. LSTMs are chosen for their ability to capture long-term dependencies in motion, while the bidirectional nature allows the network to utilize both "past" and "future" context within a window to refine estimates.

The IONet architecture uses a Bi-LSTM to process IMU sequences into polar vectors for trajectory estimation.

Instead of predicting absolute $X$ and $Y$ coordinates, the network predicts a polar vector ($\Delta l, \Delta \psi$). This decoupling of distance traveled from absolute orientation makes the learning process significantly more stable and prevents error propagation between windows. The network implicitly learns to compensate for gravity and initial velocity—factors that usually require tedious manual calibration—by recognizing patterns in how humans move.

Key Results and Evaluation

To evaluate IONet, the authors collected an extensive dataset using a Vicon optical motion capture system for millimeter-perfect ground truth. Testing involved multiple users, devices (iPhone 5, 6, and 7), and carrying modes (handheld, pocket, handbag, and trolley).

Accuracy and Robustness

IONet maintained a maximum error of approximately 2 meters within 90% of the test time, representing a 30-40% improvement over state-of-the-art PDR methods. Unlike SINS, which drifts exponentially, IONet produces stable paths even over long distances in large office environments.

Cumulative Distribution Function (CDF) of error across different phone attachments, comparing IONet to PDR and SINS.

Generalization to Non-Periodic Motion

A standout achievement of IONet is its performance in the shopping trolley experiment. While PDR failed completely due to the lack of detectable "steps," IONet's trajectory closely matched the ground truth. This demonstrates that the model learns underlying motion physics rather than just counting periodic patterns. Despite being trained only on an iPhone 7, the model performed well on older hardware, proving the learned features are related to motion physics rather than specific sensor hardware quirks.

Conclusion

IONet represents a significant shift in inertial odometry by replacing traditional integration with deep sequence learning. By treating windows independently and predicting polar displacements, it effectively "cures" the curse of drift. Future work includes developing "universal" models that automatically adapt to different phone attachments and integrating traditional physical constraints to further enhance robustness.

Summary updated

Abstract

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.
BibTeX
@inproceedings{chen2018ionet,
  author = {Changhao Chen and Xiaoxuan Lu and Andrew Markham and Niki Trigoni},
  title = {{IONet: Learning to Cure the Curse of Drift in Inertial Odometry}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2018},
  url = {https://openreview.net/forum?id=H1b_uTeuZH},
}
Figure