IMU odometry using deep learning


Creative Commons License

Oral H. K., Demir A. O., Uslu E.

V. INTERNATIONAL TURKIC WORLD CONGRESS ON SCIENCE AND ENGINEERING, Bishkek, Kyrgyzstan, 15 - 17 September 2023, pp.38-51

  • Publication Type: Conference Paper / Full Text
  • City: Bishkek
  • Country: Kyrgyzstan
  • Page Numbers: pp.38-51
  • Yıldız Technical University Affiliated: Yes

Abstract

Abstract: IMU is a measurement unit that usually consists of sensors such as
accelerometers, gyroscopes and magnetometers. Odometry is the estimation of
incremental or absolute position and orientation of the device by processing sensor
data over time. This paper focuses on how deep learning, a subtopic of artificial
intelligence, can be used for IMU odometry. IMU odometry is solely the use of data
from IMU sensors for position and orientation estimation. IMU sensor
measurements are subject to many errors and noise, and therefore the
measurements obtained may not reflect the true values. Our study filters these
measurements with deep learning algorithms to provide a more consistent
estimate. Our study uses recurrent neural networks such as LSTM, GRU and a
hybrid CNN-LSTM network to estimate the position of a micro aerial vehicle using
IMU data and compares the results with the ground-truth data by means of defined
metrics. EuRoC MAV dataset is used for this purpose. Our study also aims to
predict orientation change in terms of rotation angle using only IMU data utilizing a
deep learning model. While GRU model performed the best on position estimation
by an R2 value of 0,996, orientation estimation reached an R2 value of 0,56. The
promising results enable the use of the proposed method as an intermediary
position estimation on high level SLAM algorithms as a future work.
Keywords: IMU Odometry, Deep Learning, LSTM, GRU, CNN-LSTM.