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, including accelerometers and gyroscopes, are ubiquitous in modern mobile devices. However, their utility for indoor navigation is severely limited by the "curse of drift." Traditional Strapdown Inertial Navigation Systems (SINS) rely on the double integration of acceleration to determine position. In low-cost sensors, even minute biases or noise are amplified through this integration process, leading to cubic error growth over time. Consequently, SINS-based trajectories typically become unusable within seconds.

To circumvent this, Pedestrian Dead Reckoning (PDR) systems are often employed. PDR leverages the periodic nature of human walking to detect steps and estimate stride length. While more stable than SINS, PDR is inherently brittle. It requires extensive manual tuning for different users and walking styles, and it fails when the device's orientation relative to the body changes (e.g., moving a phone from the hand to a pocket). Most significantly, PDR is incapable of tracking non-periodic motions, such as pushing a shopping cart or a stroller, where no distinct "steps" occur.

The IONet Framework

IONet addresses these challenges by reframing inertial odometry as a sequence-to-sequence learning problem. The core innovation is "breaking the cycle" of continuous integration by segmenting raw IMU data into independent, overlapping windows of approximately 2 seconds (200 frames).

Physics-Inspired Deep Learning

Instead of a purely black-box approach, IONet is grounded in Newtonian mechanics. The authors derive a model that calculates displacement in a local coordinate frame. By predicting motion in polar coordinates—specifically the change in distance ($\Delta l$) and change in heading ($\Delta \psi$)—the model decouples the estimated displacement from the device's initial unknown orientation.

The IONet architecture uses a Bidirectional LSTM to process IMU sequences into polar vectors, which are then aggregated into a full pose sequence.

The architecture employs a 2-layer Bidirectional Long Short-Term Memory (Bi-LSTM) network. The bidirectional nature allows the model to utilize both past and future context within a window to interpret complex motion patterns. Crucially, hidden states are not passed between windows. This independence ensures that errors do not accumulate across the entire trajectory, effectively neutralizing the drift that plagues traditional integration methods.

Experimental Evaluation

To train and validate IONet, a comprehensive dataset was collected using a Vicon motion capture system for millimeter-accurate ground truth. The data spanned various device attachments (handheld, pocket, handbag) and multiple users to ensure robustness.

Key Results and Generalization

IONet consistently outperformed both SINS and PDR across all experimental conditions. In standard walking tests, IONet achieved a 30-40% improvement in accuracy over state-of-the-art PDR systems, with maximum errors remaining around 2 meters at a 90% confidence interval.

Comparison of tracking performance for a phone in a pocket. IONet (green) closely follows the ground truth (Tango, blue), while SINS (purple) drifts immediately and PDR (red) accumulates significant error.

One of the most significant achievements of IONet is its ability to generalize to non-periodic motion. In tests involving a shopping trolley, IONet maintained accuracy comparable to high-end visual-inertial systems like Google Tango, despite using only IMU data. This "step-less" advantage allows IONet to track any smooth motion that PDR-based systems simply cannot detect.

Conclusion and Future Work

IONet demonstrates that deep recurrent neural networks can effectively "learn" to interpret noisy inertial data, replacing error-prone mathematical integration with robust pattern recognition. While the system shows remarkable stability, some heading drift remains over very long trajectories, and performance can degrade if sensor bias significantly exceeds training parameters.

Future research will focus on developing unified models capable of automatically detecting device attachment types and extending the framework to more diverse platforms, including autonomous robots and drones, where non-periodic motion is the standard.

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