Skip to Main Content
 

Global Search Box

 
 
 
 

Files

File List


LEE_Thesis.pdf (11.02 MB)

ETD Abstract Container

Abstract Header

Deep Learning for Visual-Inertial Odometry: Estimation of Monocular Camera Ego-Motion and its Uncertainty

Lee, Hong Yun
http://rave.ohiolink.edu/etdc/view?acc_num=osu156331321922759

Abstract Details

2019, Master of Science, Ohio State University, Aero/Astro Engineering.
For autonomous robotic operations in GPS-denied environments, visual odometry (VO) or visual-inertial odometry (VIO) provides a means of estimating velocity and position for autonomous vehicles. Since these vision-based solutions rely on cameras, the calibration for the camera is a necessary step as the raw camera lens has uncalibrated distortion. With the recent advance of deep learning (DL), it is shown that DL could learn entire VO pipeline so that the designers do not need to consider the cumbersome parameter tuning for the VO algorithm. However, the DL-based method requires heavy computational memory. Usually, the conventional VO/VIO algorithms attempted to reduce the computational power for embedded computers such as smartphones or onboard computers typically used in unmanned aerial vehicles. Likewise, this work attempts to reduce the computational burden of DL-based VIO problem. The assumed system has a gyroscope, an accelerometer, and an onboard computer with limited memory capacity. This work proposes a relatively shallow convolutional neural net (CNN) structure that process images to estimate the translational and the rotational motion of a monocular camera, also known as the ego-motion, and its uncertainty. Using special Euclidian group and its underlying algebra, the rotational information from the gyroscope can be combined with other sensors to provide the necessary aiding information to estimate the ego-motion of the camera. The VO algorithm needs to estimate the translational velocity and position with a minimal error. Therefore, an embedded linear Kalman filter is used to correct the velocity estimates based on acceleration information from an accelerometer. In this case, neural nets are used to optimize the Kalman filter process noise covariance for the prediction step of the Kalman filter. The proposed architecture in this work has reduced, by more than 50%, of the memory required compared to the previously proposed DL architecture. The performance is measured with the root mean square error (RMSE). The overall positional RMSE of the architecture is about 4.2 meters per 100 meters. The overall RMSE for the displacement is 2.97 meters. The RMSE per 500 meters for KITTI dataset sequence 10 is 14.38 m and 7.78 m before and after the correction respectively.
James Gregory (Advisor)
Matthew McCrink (Committee Member)
Alper Yilmaz (Committee Member)
139 p.
Aerospace Engineering

Recommended Citations

Citations

  • Lee, H. Y. (2019). Deep Learning for Visual-Inertial Odometry: Estimation of Monocular Camera Ego-Motion and its Uncertainty [Master's thesis, Ohio State University]. OhioLINK Electronic Theses and Dissertations Center. http://rave.ohiolink.edu/etdc/view?acc_num=osu156331321922759

    APA Style (7th edition)

  • Lee, Hong. Deep Learning for Visual-Inertial Odometry: Estimation of Monocular Camera Ego-Motion and its Uncertainty . 2019. Ohio State University, Master's thesis. OhioLINK Electronic Theses and Dissertations Center, http://rave.ohiolink.edu/etdc/view?acc_num=osu156331321922759.

    MLA Style (8th edition)

  • Lee, Hong. "Deep Learning for Visual-Inertial Odometry: Estimation of Monocular Camera Ego-Motion and its Uncertainty ." Master's thesis, Ohio State University, 2019. http://rave.ohiolink.edu/etdc/view?acc_num=osu156331321922759

    Chicago Manual of Style (17th edition)

osu156331321922759
1,699
© 2019, all rights reserved.
This open access ETD is published by The Ohio State University and OhioLINK.