Deep Learning for Inertial Navigation

Original article was published by Barak Or on Deep Learning on Medium


Deep Learning for Inertial Navigation

A short review of cutting edge deep learning-based solutions for inertial navigation.

Barak et al 2020

Introduction

Many vision aiding navigation approaches were presented in the last decade as there is a wide range of applications these days (Huang, 2019). Saying that the classical field of inertial navigation with low-cost inertial sensors as the only source of information is starting getting attention with novel deep learning methods to involve in it.

In this post, we review the integration of deep learning in classical Inertial Navigation System (INS) with Inertial Measurement Units (IMU’s) only. First, we present some cutting edge architectures for improved speed estimation, noise reduction, zero-velocity detection, and attitude & position prediction. Secondly, the KITTI and OxIOD dataset are discussed. Lastly, schemes of pedestrian inertial navigation with deep learning are presented.

Cutting edge deep learning-based Solutions

One of the main problems in the navigation field is speed estimation. As the estimation becomes accurate it affect also on the position solution. In a work published in 2018 by Cortes et al, a deep learning-based speed estimation approach was suggested. The main idea was to add a speed constrain to the classical Inertial Navigation System (INS). They estimated the speed from the IMU only by using a CNN and then constrained the INS solution by this prediction. Formulating this estimation as a regression deep learning task, where the inputs are the six-channel of the IMU over a few seconds, and the output is the speed, leads to improved trajectory tracking as also motion mode classification.

Cortes et al, 2018

The next work I want to present regarding noise reduction. As many low-cost sensors suffer from the high magnitude of noise and characterized by a noise profile, where the noise changes with time, there is a need to filter it. But as these noise profiles are difficult to estimate, using a deep learning-based approach seems to solve this issue. Chen et al (2018) presented a novel deep learning approach to deal with many error sources in the sensor signals. By doing that, the sensors’ signals can be corrected and only then to be used in the navigation scheme. They reported 80% accuracy on the correct identification of the IMU signals. CNN was also used in this work, where it includes 5 convolutional layers and one fully-connected layer.

Chen et al, 2018

Another work, by Wagstaff and Kelly (2018), deals with indoor navigation tasks where a scheme to detect foot zero-velocity is presented. By doing that, the accuracy of the velocity estimation is provided, and by the general INS, accuracy is improved. The detection was done by designing a Long Short-Term Memory (LSTM) neural network. By evaluating the designed scheme for more than 7.5 [km] of indoor pedestrian locomotion data, they reported a reduction of over 34% positioning error. Their architecture includes 6-layers LSTM with 80 units per each one of them and a single fully-connected layer after the LSTM.

Wagstaff and Kelly (2018)

The last work I want to discuss is related to one of the main problems in the navigation field: attitude and position prediction. Achieving precise state estimation of attitude is very important to multirotor systems, as small errors might lead to instability and end in disaster. A work by Al-Sharman et al (2019) presents a deep learning-based state estimation enhancement with a particular application to attitude estimation. They tackled the problem of precise attitude estimation by noise reduction technique, where they identified the measurement noise characteristics. They used a simple multilayer neural network with a dropout technique and exhibited superiority over the conventional approaches.

Al-Sharman et al (2020)

Common dataset

To evaluate the different suggestion approaches, there are two common datasets to work with, the first one is the KITTI dataset of Karlsruhe Institute of Technology. It contains a big amount of data: Velodyne, IMU, GPS, camera calibration, grayscale stereo sequences, 3D object trucklet labels, and many more. The paper by Geiger et al (2013) reviews the entire dataset. Alto KITTI is the main dataset in the field, I want to mention additional pretty new dataset called OxIOD by Oxford. It is used for deep inertial odometry and the entire information is available in the paper by Chen et al (2018).

Geiger et al (2013)
Chen et al (2018)

Pedestrian Inertial Navigation

Recently, there is a growing interest in applying deep learning techniques to motion sensing and location estimation of pedestrians. In the work of Chen et al (2020) the methods, data set and on-device interface of deep learning-based pedestrian INS are greatly reviewed. In their work, they proposed the L-IONet, a framework to learn the inertial tracking from raw data. The architecture contains mainly 1d dilated convolutional layers, where they inspired by WaveNet, and fully-connected layers.

Another work, by Klein et-al (2020) presents the StepNet: a deep learning approach for step length estimation. The authors addressed the pedestrian indoor dead reckoning by a family of deep learning-based approaches to regress the step-length. The suggested StepNet outperforms traditional approaches as described in their paper.

Klein et al (2020)

Summary

With the high popularity of deep learning keeps rising, it appears to solve many classical problems in the field of inertial navigation. As we described in this post, some researches addressing the integration of deep learning and inertial navigation with promise results.

References

Huang, Guoquan. “Visual-inertial navigation: A concise review.” 2019 International Conference on Robotics and Automation (ICRA). IEEE, 2019.

Cortés, Santiago, Arno Solin, and Juho Kannala. “Deep learning based speed estimation for constraining strapdown inertial navigation on smartphones.” 2018 IEEE 28th International Workshop on Machine Learning for Signal Processing (MLSP). IEEE, 2018.

Chen, Hua, et al. “Improving inertial sensor by reducing errors using deep learning methodology.” NAECON 2018-IEEE National Aerospace and Electronics Conference. IEEE, 2018.‏

Wagstaff, Brandon, and Jonathan Kelly. “LSTM-based zero-velocity detection for robust inertial navigation.” 2018 International Conference on Indoor Positioning and Indoor Navigation (IPIN). IEEE, 2018.

Al-Sharman, Mohammad K., et al. “Deep-learning-based neural network training for state estimation enhancement: application to attitude estimation.” IEEE Transactions on Instrumentation and Measurement 69.1 (2019): 24–34.

Geiger, Andreas, et al. “Vision meets robotics: The kitti dataset.” The International Journal of Robotics Research 32.11 (2013): 1231–1237.

Chen, Changhao, et al. “Oxiod: The dataset for deep inertial odometry.” arXiv preprint arXiv:1809.07491 (2018).

Chen, Changhao, et al. “Deep-Learning-Based Pedestrian Inertial Navigation: Methods, Data Set, and On-Device Inference.” IEEE Internet of Things Journal 7.5 (2020): 4431–4441.

Klein, Itzik, and Omri Asraf. “StepNet — Deep Learning Approaches for Step Length Estimation.” IEEE Access 8 (2020): 85706–85713.‏