IONet: Learning to Cure the Curse of Drift in Inertial Odometry (1802.02209v1)

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.

• The paper's main contribution is the development of IONet, a deep neural network that treats inertial odometry as a sequential learning problem to mitigate drift.
• The study employs a two-layer bidirectional LSTM to transform raw IMU data into polar displacement vectors, enhancing robustness to sensor placement.
• Extensive experiments show IONet outperforms traditional methods in diverse indoor environments and under irregular motion conditions.

Analysis of IONet: A Deep Learning Approach to Inertial Odometry

The paper presents an innovative approach to address the challenge of drift correction in inertial odometry using deep learning, introducing the framework known as IONet. Inertial odometry is crucial for various applications in indoor localization, yet low-cost inertial sensors suffer from significant drift issues due to noise and bias. This leads to the rapid degradation of odometry accuracy as measurements are integrated over time.

Overview and Contributions

IONet proposes a shift from traditional models by treating inertial tracking as a sequential learning problem, thus forgoing the dependency on hand-engineered, step-based pedestrian dead reckoning (PDR) methods. The proposed framework includes:

1. Conceptualizing inertial tracking as a sequential learning problem via a novel sequence-based model derived from Newtonian mechanics.
2. Developing a deep neural network to learn location transformations from raw inertial measurement unit (IMU) data within a polar coordinate framework, making the model robust to different IMU attachment configurations.
3. Demonstrating the model's effectiveness through extensive empirical evaluations, showcasing its ability to generalize across a variety of environments and motion types, including non-periodic motions like those of a shopping trolley.

Methodology

The paper innovatively addresses the drift problem by segmenting inertial data into independent windows to reduce continuous error propagation. It employs a novel approach where the change in navigation state over each window is accumulated through simple summation. This is akin to resetting an integrator, a strategy often used in control systems to manage error degradation.

The neural network architecture is based on deep recurrent neural networks (RNNs) and specifically on a two-layer bidirectional long short-term memory (LSTM) model. This design effectively captures temporal dependencies within the IMU data, allowing the network to learn the transformation from raw data to polar displacement vectors.

Experimental Results

The authors conducted experiments involving different users and devices, along with tests in large-scale indoor environments, demonstrating superior performance relative to traditional algorithms like PDR and strapdown inertial navigation systems (SINS). Remarkably, IONet sustained lower error rates and could adapt to various motion patterns and sensor placements. In particular, the model achieved commendable results in scenarios with irregular motion patterns, such as trolley movements, outperforming a state-of-the-art visual-inertial odometry system under certain conditions.

Implications and Future Directions

IONet represents a significant step forward in inertial navigation by using a data-driven method to bypass the limitations inherent in classical algorithms. Its flexibility in adapting to different sensor configurations and its robustness across a range of motion types offer promising potential for real-world applications in personal navigation systems and robotics.

For future developments, challenges remain in handling high-bias IMU data and ensuring model robustness across diverse user behaviors and environments. Enhancing the capability of a singular model to manage various attachments is another intriguing research avenue. Furthermore, integrating sequence-based physical models with data-driven approaches like IONet could foster advancements in generalization and performance.

In conclusion, IONet provides a novel and effective methodology for tackling drift challenges in inertial odometry, offering new opportunities for enhancing indoor localization technologies. As AI and sensor technologies progress, the integration of deep learning with traditional physics-based models may continue to alleviate long-standing challenges in this domain.