Efficient and accurate stereo matching for cloth manipulation

Xu, Tian (2016) Efficient and accurate stereo matching for cloth manipulation. PhD thesis, University of Glasgow.

Full text available as:
[img]
Preview
PDF
Download (21MB) | Preview

Abstract

Due to the recent development of robotic techniques, researching robots that can assist in everyday household tasks, especially robotic cloth manipulation has become popular in recent years. Stereo matching forms a crucial part of the robotic vision and aims to derive depth information from image pairs captured by the stereo cameras. Although stereo robotic vision is widely adopted for cloth manipulation robots in the research community, this remains a challenging research task. Robotic vision requires very accurate depth output in a relatively short timespan in order to successfully perform cloth manipulation in real-time. In this thesis, we mainly aim to develop a robotic stereo matching based vision system that is both efficient and effective for the task of robotic cloth manipulation. Effectiveness refers to the accuracy of the depth map generated from the stereo matching algorithms for the robot to grasp the required details to achieve the given task on cloth materials while efficiency emphasizes the required time for the stereo matching to process the images. With respect to efficiency, firstly, by exploring a variety of different hardware architectures such as multi-core CPU and graphic processors (GPU) to accelerate stereo matching, we demonstrate that the parallelised stereo-matching algorithm can be significantly accelerated, achieving 12X and 176X speed-ups respectively for multi-core CPU and GPU, compared with SISD (Single Instruction, Single Data) single-thread CPU. In terms of effectiveness, due to the fact that there are no cloth based testbeds with depth map ground-truths for evaluating the accuracy of stereo matching performance in this context, we created five different testbeds to facilitate evaluation of stereo matching in the context of cloth manipulation. In addition, we adapted a guided filtering algorithm into a pyramidical stereo matching framework that works directly for unrectified images, and evaluate its accuracy utilizing the created cloth testbeds. We demonstrate that our proposed approach is not only efficient, but also accurate and suits well to the characteristics of the task of cloth manipulations. This also shows that rather than relying on image rectification, directly applying stereo matching to unrectified images is effective and efficient. Finally, we further explore whether we can improve efficiency while maintaining reasonable accuracy for robotic cloth manipulations (i.e.~trading off accuracy for efficiency). We use a foveated matching algorithm, inspired by biological vision systems, and found that it is effective in trading off accuracy for efficiency, achieving almost the same level of accuracy for both cloth grasping and flattening tasks with two to three fold acceleration. We also demonstrate that with the robot we can use machine learning techniques to predict the optimal foveation level in order to accomplish the robotic cloth manipulation tasks successfully and much more efficiently. To summarize, in this thesis, we extensively study stereo matching, contributing to the long-term goal of developing effective ways for efficient whilst accurate robotic stereo matching for cloth manipulation.

Item Type: Thesis (PhD)
Qualification Level: Doctoral
Keywords: Stereo matching, GPU, multi-core CPU, parallel processing, robotic vision, cloth manipulation.
Subjects: Q Science > QA Mathematics > QA75 Electronic computers. Computer science
Colleges/Schools: College of Science and Engineering > School of Computing Science
Funder's Name: UNSPECIFIED
Supervisor's Name: CockShott, Dr. Paul
Date of Award: 2016
Depositing User: Dr Tian Xu
Unique ID: glathesis:2016-7262
Copyright: Copyright of this thesis is held by the author.
Date Deposited: 26 Apr 2016 13:18
Last Modified: 16 May 2016 10:03
URI: http://theses.gla.ac.uk/id/eprint/7262

Actions (login required)

View Item View Item

Downloads

Downloads per month over past year