README.md

IMU Debiasing

Overview

This project focuses on debiasing (denoising) Inertial Measurement Unit (IMU) data, including gyroscope and accelerometer measurements. A neural network explicitly models the bias dynamics, while a neural ODE on $SO(3)$ is designed for training. The loss is computed using ground truth orientation, velocity, and position, without requiring ground truth for bias.

The overall framework is illustrated below:

Figure 1: The hierarchical neural ODE framework.

Paper

Coming soon

Setup

Create a Virtual Environment

To ensure an isolated environment for dependencies, run the following commands (Python version: 3.10.12):

python3 -m venv venv
source venv/bin/activate
pip install -r Requirments.txt

Prepare the Dataset

The required IMU data is provided in the data/ folder. Alternatively, you can download data from the following sources:

• EUROC
• TUM-VI

Note: For TUM-VI, the IMU data is extracted from the raw-data rosbag, not the calibrated rosbag, but uses the synthesized timestamps from the calibrated rosbag.

Begin training!

Run the following commands in the terminal to start training (ensure you are using python interpreter from venv/)

python3 BiasDy/mainEuroc.py
python3 BiasDy/mainTUM.py
python3 BiasDy/mainFetch.py

Results

The full results will be saved to ./results. Partial results:

Figure 2: The Euler angles results for MH_04_difficult.

Figure 3: The Euler angles results for dataset_room4.

Citation

To be added once the corresponding paper is published.

Acknowledgments

This project incorporates code and ideas from the following sources:

• denoise-imu-gyro by M. Brossard
• torchdiffeq by Ricky T. Q. Chen
• NeuralCDE by Patrick Kidger