Scale and Direction Guided GAN for Inertial Sensor Signal Enhancement

Yifeng Wang, Yi Zhao, Jiye Liang, Jianqing Liang

Abstract

Low-cost Inertial Measurement Units (IMUs) are ubiquitous in consumer electronics but suffer from significant intrinsic errors like quantization noise and thermal drift. These errors accumulate over time, hindering precise applications like 3D motion tracking. Traditional deep learning solutions often require paired "clean" and "noisy" training data, which is logistically difficult to collect for IMUs. This paper, presented at IJCAI-24, introduces SDG-GAN, a novel generative model that enhances low-cost inertial signals without requiring paired data. By employing two novel guidance mechanisms—multi-scale "zoom" consistency (SG-GAN) and dynamic auxiliary task weighting (DG-GAN)—the model achieves state-of-the-art results in signal denoising and trajectory reconstruction. Experimental results across 15 different smartphones demonstrate a 97% noise elimination rate and significant improvements in downstream task accuracy.

Introduction

Inertial Measurement Units (IMUs), comprising accelerometers and gyroscopes, are essential for modern navigation and motion tracking. However, consumer-grade IMUs are plagued by noise that makes long-term precision nearly impossible with raw data. While deep learning offers a solution, most high-performing models rely on perfectly aligned paired datasets. SDG-GAN addresses this bottleneck by leveraging unsupervised and weakly-supervised learning strategies to "reimagine" high-quality signals from noisy inputs.

Methodology

The SDG-GAN framework operates through two distinct guidance components tailored to different data availability scenarios.

Scale-Guided GAN (SG-GAN)

In fully unsupervised settings, the model utilizes a self-supervised zoom constraint. The generator processes signals at multiple resolutions (e.g., 1.0x, 1.2x, 2.0x). A "zoom loss" forces the generator to ensure that a high-resolution signal, when downsampled, matches the signal generated at a lower resolution. This prevents the model from generating artifacts and ensures structural integrity across different scales.

The SDG-GAN architecture utilizes multi-scale zoom consistency to ensure signal integrity across different resolutions without paired training data.

Direction-Guided GAN (DG-GAN)

When some auxiliary information is available, the model incorporates tasks like Attitude Estimation (AE) and Displacement Prediction (DP). To prevent these auxiliary tasks from conflicting with the primary denoising goal, a Learning Direction Guidance mechanism is introduced. This mechanism monitors feature consistency and dynamically suppresses the influence of any auxiliary task that pulls the model in a direction contradicting the main generation task.

Experimental Results

The authors evaluated SDG-GAN using a comprehensive setup involving 15 different smartphones and high-precision ground truth equipment, including a robotic arm and an optical motion capture system.

Performance and Generalization

SDG-GAN achieved an average noise elimination rate of over 97%. It reduced attitude estimation error from ~10.7° to 1.85° and position error from 1.09m to 0.14m. Notably, the model demonstrated excellent generalization; a model trained on a single phone type performed effectively across 14 other hardware configurations without retraining.

Trajectory Reconstruction

In practical tests such as "writing in the air," SDG-GAN was the only method capable of producing a morphologically accurate 3D path from a single low-cost sensor, effectively repairing intrinsic features that standard filters often miss.

3D trajectory reconstruction of a complex motion path, showing the high fidelity achieved by the optimized model in practical tracking scenarios.

Conclusion

SDG-GAN represents a significant leap in inertial sensor signal processing by overcoming the need for paired training data. The combination of multi-scale consistency and dynamic task weighting allows for robust denoising and accurate motion reconstruction. Future work will explore reducing computational overhead for real-time deployment on low-power microcontrollers and testing in more irregular, high-impact environments.

Summary updated

BibTeX
@inproceedings{wang2024scale,
  author = {Yifeng Wang and Yi Zhao and Jiye Liang and Jianqing Liang},
  title = {{Scale and Direction Guided GAN for Inertial Sensor Signal Enhancement}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2024},
  url = {https://openreview.net/forum?id=mJJF7Ea1NZ},
}
Figure