University of Illinois Urbana-Champaign

Improved pose estimation accuracy of monocular deep visual odometry against dynamic entities via adversarial training

Chen, Jushan

Content Files
CHEN-THESIS-2024.pdf

Permalink

https://hdl.handle.net/2142/124418

Description

Title
Improved pose estimation accuracy of monocular deep visual odometry against dynamic entities via adversarial training
Author(s)
Chen, Jushan
Issue Date
2024-05-03
Director of Research (if dissertation) or Advisor (if thesis)
Mehr, Negar
Department of Study
Aerospace Engineering
Discipline
Aerospace Engineering
Degree Granting Institution
University of Illinois at Urbana-Champaign
Degree Name
M.S.
Degree Level
Thesis
Keyword(s)
  • robust visual odometry
  • adversarial training
Abstract
Camera pose estimation, also known as ego-motion estimation, is an important problem in visual odometry, which relies on visual cues to estimate the camera pose. Existing deep learning-based visual odometry frameworks have started to outperform the camera pose prediction accuracy of geometry-based visual odometry methods in some benchmarks. However, it remains challenging to accurately estimate the ego-motion in dynamic scenes because visual odometry frameworks typically rely on a static scene assumption. To address this issue, previous works have improved the robustness of visual odometry networks to dynamic scenes using highly complex architectures which make them time-consuming to train. Instead, in this thesis, we propose a lightweight robust visual odometry network via a novel adversarial training scheme. Our adversarial training scheme simulates the motion of dynamic obstacles by creating a sequence of artificially altered trajectories of dynamic entities. This scheme is a data augmentation framework that can be easily incorporated into an existing deep visual odometry model without modifying its internal modules. We evaluate the performance of our robust visual odometry model on selected sequences of the KITTI odometry dataset containing a large number of dynamic objects. We show that the pose estimation accuracy of our robust visual odometry network outperforms our backbone deep visual odometry network by 64% on average.
Graduation Semester
2024-05
Type of Resource
Thesis
Handle URL
https://hdl.handle.net/2142/124418
Copyright and License Information
Copyright 2024 Jushan Chen

Owning Collections

Graduate Dissertations and Theses at Illinois PRIMARY
Graduate Theses and Dissertations at Illinois