Visual 6D object pose estimation and tracking

Zheng, Linfang ORCID: 0009-0001-8006-8325 (2024). Visual 6D object pose estimation and tracking. University of Birmingham. Ph.D.

Preview

Zheng2024PhD.pdf Text - Accepted Version Available under License All rights reserved. Download (62MB) | Preview

Abstract

Visual 6D object pose estimation and tracking is crucial for enabling machines to understand and interact with the three-dimensional world. Despite significant advancements, challenges persist in this field. Existing methods often struggle in complex real-world scenarios, particularly with symmetric and textureless objects under occlusion. Category-level methods also face limitations in generalizability, especially with complex-shaped objects and noisy environments. Additionally, there is a significant gap in effective methods for category-level object pose refinement, which is crucial for achieving high-precision pose information with previously unseen objects. To address these challenges, this thesis proposes three approaches. First, an instance-level object pose tracking method is introduced, leveraging temporal information with augmented autoencoder-based reconstruction to enhance robustness to symmetric and textureless objects under occlusion. Second, a hybrid scope feature extraction layer (HS-layer) is presented for category-level object pose estimation. The HS layer encodes translation and scale information, perceives geometric structural information for handling complex-shaped objects with robustness to outliers. Lastly, a method that combines latent geometric feature extraction and learnable affine transformation is proposed to address shape discrepancy issues in category-level object pose refinement, improving pose refinement accuracy and generalizability. Extensive experiments validate the effectiveness of these approaches in advancing practical applications of visual 6D object pose estimation and tracking. Additionally, all the proposed methods exhibit real-time performance, crucial for real-life applications.

Type of Work:	Thesis (Doctorates > Ph.D.)
Award Type:	Doctorates > Ph.D.
Supervisor(s):	Supervisor(s) Email ORCID Chang, Hyung Jin UNSPECIFIED UNSPECIFIED Leonardis, Ales UNSPECIFIED UNSPECIFIED
Licence:	All rights reserved
College/Faculty:	Colleges > College of Engineering & Physical Sciences
School or Department:	School of Computer Science
Funders:	Other
Other Funders:	National Natural Science Foundation of China under Grant No. 62073159 and Grant No. 62003155, Korea government (MSIT), the Shenzhen Key Laboratory of Control Theory and Intelligent Systems
Subjects:	T Technology > T Technology (General)
URI:	http://etheses.bham.ac.uk/id/eprint/15459

Actions

	Request a Correction
	View Item

Downloads