Deep Inertial Odometry for Robust Pedestrian Navigation

This direction is merged into another direction: inertial sensor drift. Its last version stays readable and citable here.

Version 1 of 124 papers27 conceptsfrozen 2026-10-04

Traditional indoor navigation has long relied on a simple heuristic: if you can count a person's steps and guess how long each step is, you can track where they are going. This approach, known as Pedestrian Dead Reckoning (PDR), works well in a laboratory but often fails in the messy reality of daily life. Recent research is shifting away from these rigid step-counting models toward Deep Inertial Odometry, which uses neural networks to learn the underlying physics of motion directly from raw sensor data.

The Brittle Nature of Step Counting

The fundamental flaw in traditional PDR is its reliance on "step-and-stride" assumptions. Most systems use a fixed constant to estimate distance based on step count, but as Shangsheng Wen and colleagues point out in Smartphone-Based Pedestrian Dead Reckoning Integrated with Data-Fusion-Adopted Visible Light Positioning, these constants fail to account for individual gait differences or changes in walking speed.

Even when systems try to be more dynamic, they remain sensitive to "scale errors." Pierre-Yves Lajoie and the team behind PEOPLEx note that without an external reference, motion models often produce trajectories with incorrect scaling because step lengths vary wildly across different users and environments. Furthermore, Boqi Wu and Ying Chi observe that simple threshold-based methods often struggle to even identify the "zero-velocity" moments (when a foot touches the ground) necessary to reset sensor drift.

When Steps Disappear

The most significant limitation of step-based models is that they only work when a person is actually taking steps. In Inertial Odometry on Handheld Smartphones, Arno Solin and his co-authors argue that this dependency prevents accurate tracking in "generalized mobility" scenarios. If a user is pushing a stroller, riding an elevator, or moving a shopping trolley, there are no rhythmic footfalls to detect.

To solve this, researchers like Changhao Chen and Andrew Markham have pioneered "motion-agnostic" models. By creating OxIOD: The Dataset for Deep Inertial Odometry, they provided the first large-scale data for training models that don't care how you move. Their framework, L-IONet, uses dilated causal convolutions to learn displacement directly from the Inertial Measurement Unit (IMU) data, allowing a phone to track movement even when it is sitting on a trolley or buried in a handbag.

Fusing Physics with Deep Learning

While deep learning is powerful at recognizing patterns, it can sometimes lack the mathematical rigor of classical physics. The cutting edge of this field involves hybrid models that combine the two.

In Invariant Extended Kalman Filtering for Pedestrian Deep-Inertial Odometry, Shiyu Bai and Weisong Wen demonstrate a system where a neural network estimates forward speed while an Invariant Extended Kalman Filter (IEKF) handles the geometric orientation and sensor bias. This "Deep-Inertial" approach provides a safety net: the neural network handles the complex, non-linear reality of human movement, while the filter ensures the trajectory remains physically consistent.

Opportunistic and Plug-and-Play Systems

The ultimate goal of this research direction is "opportunistic" positioning—systems that work anywhere without prior setup. PEOPLEx represents this trend by using a nonlinear factor graph to fuse IMU data with whatever signals happen to be available, such as Wi-Fi, Bluetooth (BLE), or Ultra-Wideband (UWB).

By moving away from the rigid assumption that "walking equals stepping," these new models transform the smartphone from a simple pedometer into a sophisticated inertial navigation system capable of decimeter-level accuracy in the most unpredictable real-world environments.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_deep_inertial_odometry_for_robust_pedest_449f7dc9,
  title = {Deep Inertial Odometry for Robust Pedestrian Navigation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/deep-inertial-odometry-for-robust-pedestrian-navigation/rdr_449f7dc9b49d4ee7a61988137b2e203e},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_deep_inertial_odometry_for_robust_pedest_449f7dc9_v1,
  title = {Deep Inertial Odometry for Robust Pedestrian Navigation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/deep-inertial-odometry-for-robust-pedestrian-navigation/rdr_449f7dc9b49d4ee7a61988137b2e203e/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}