Neural Inertial Localization

Sachini Herath, David Caruso, Chen Liu, Yufan Chen, Yasutaka Furukawa

Indoor localization remains a significant challenge in mobile computing. While GPS provides reliable positioning outdoors, indoor environments typically necessitate a patchwork of WiFi signals, Bluetooth beacons, or privacy-invasive camera systems. Researchers from Simon Fraser University and Meta Reality Labs have introduced NILoc, a novel deep learning framework that enables absolute indoor localization using only raw Inertial Measurement Unit (IMU) data from standard smartphone sensors.

The Challenge of Inertial Localization

Most existing indoor localization systems come with significant trade-offs. WiFi-based systems are dependent on infrastructure density, while visual systems consume high power and raise privacy concerns. Pure inertial navigation, or "dead reckoning," tracks relative motion but suffers from cumulative sensor errors known as "drift," which quickly renders position estimates inaccurate.

NILoc addresses this by asking: Can we map a specific motion history to a specific location? In a structured indoor environment, certain sequences of movement—such as walking a specific distance, turning, and entering a room—are only physically possible in a few locations. NILoc learns these scene-specific "motion signatures" to provide absolute coordinates using energy-efficient, privacy-preserving sensors.

The NILoc Framework

The NILoc architecture processes raw sensor data through a multi-stage neural pipeline to estimate position without requiring external signals or a floorplan at inference time.

Overview of the NILoc pipeline: Raw IMU data is processed into velocity sequences, which are then mapped to location likelihoods and final estimates.
  1. Neural Inertial Navigation: Instead of regressing location directly from noisy raw data, the system first utilizes a neural inertial navigation technique (RoNIN) to convert IMU history into a sequence of 2D velocity vectors. This provides a cleaner intermediate representation of the user's motion.
  2. Two-Branch Transformer: The core of the system is a scene-specific network featuring two branches:
  3. Velocity Branch: A Temporal Convolutional Network (TCN) compresses long motion histories, followed by a Transformer encoder. A translation-aware decoder then maps these features to a 2D likelihood map.
  4. Auto-regressive Location Branch: This branch ensures temporal consistency by combining current velocity features with previous location estimates, effectively filtering predictions to ensure the estimated path is physically plausible.

To overcome the data-hungry nature of Transformers, the authors developed a synthetic data generation pipeline, simulating millions of realistic walking trajectories based on existing floorplans to augment their training sets.

Dataset and Benchmarking

The researchers introduced a massive new benchmark for inertial localization, comprising 53 hours of data collected across three distinct buildings. The ground truth was established using high-accuracy Visual-Inertial SLAM, capturing "natural" phone handling behaviors such as keeping the device in a pocket, hand, or during active calls.

Visualization of the extensive trajectory dataset collected in Office C, covering over 21 km of movement.

Key Results and Performance

NILoc was evaluated against traditional baselines, including Particle Filters (PF) and state-of-the-art map-matching via Conditional Random Fields (CRF).

  • Accuracy: NILoc frequently outperformed baselines that had explicit access to floorplans. In one test environment ("Office C"), NILoc achieved a 72.9% success rate within a 1-meter threshold, significantly higher than the 44.3% achieved by the CRF baseline.
  • Efficiency: By avoiding the computationally expensive dynamic programming required by CRF methods, NILoc runs 20–30 times faster, making it suitable for real-time applications on mobile or AR devices.
  • Implicit Constraints: A key insight of the work is that NILoc learns environmental constraints (like walls and corridors) implicitly during training. Consequently, it does not require a floorplan image during inference.

Limitations and Future Directions

While NILoc represents a breakthrough in floorplan-free localization, it faces challenges in highly symmetrical environments or large open spaces (like atriums) where motion signatures are less distinct. Currently, the model must be trained for a specific building. Future research will focus on generalized models and rapid adaptation techniques for new environments.

For more details, the authors have released their code and dataset at sachini.github.io/niloc.

Summary updated

Abstract

This paper proposes the inertial localization problem, the task of estimating the absolute location from a sequence of inertial sensor measurements. This is an exciting and unexplored area of indoor localization research, where we present a rich dataset with 53 hours of inertial sensor data and the associated ground truth locations. We developed a solution, dubbed neural inertial localization (NILoc) which 1) uses a neural inertial navigation technique to turn inertial sensor history to a sequence of velocity vectors; then 2) employs a transformer-based neural architecture to find the device location from the sequence of velocities. We only use an IMU sensor, which is energy efficient and privacy preserving compared to WiFi, cameras, and other data sources. Our approach is significantly faster and achieves competitive results even compared with state-of-the-art methods that require a floorplan and run 20 to 30 times slower. We share our code, model and data at https://sachini.github.io/niloc.
BibTeX
@inproceedings{herath2022neural,
  author = {Sachini Herath and David Caruso and Chen Liu and Yufan Chen and Yasutaka Furukawa},
  title = {{Neural Inertial Localization}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2022},
  doi = {10.1109/cvpr52688.2022.00649},
}
Figure