Dissertation/Thesis Abstract

Exploiting Motion Information for Video Analysis in Sequences With Moving Camera
by Arvanitidou, Marina Georgia, Eng.D., Technische Universitaet Berlin (Germany), 2015, 184; 10697774
Abstract (Summary)

The processing of low-level information towards the extraction of high-level information, and specifically object segmentation, constitutes a great challenge for image processing and computer vision. With this respect, motion estimation is a task of major importance, since motion constitutes arguably one of the most valuable underlying clues in image sequences. Furthermore, motion is shown to be strongly connected with the human visual system. Further possibilities are thus emerging in the field of visual quality assessment for developing appropriate motion exploitation strategies that are aligned with the human visual system. This thesis focuses on exploiting motion for video analysis in image sequences captured by a moving camera and provides an appropriate evaluation framework. Firstly, motion induced by the camera movement has to be distinguished from motion resulting from the moving content itself. Therefore, the first part of the thesis is devoted to global motion estimation, i.e. the estimation of background motion. Outlier regression techniques are employed for the formulation of parametric models for global motion. It is shown that this modelling benefits from the consideration of block information since it implicitly contains information regarding foreground objects that move independently of the background region. Moreover, the parametric modelling of global motion is shown to have a positive influence towards enhancing conventional motion prediction. The second part of the thesis deals with object segmentation. A short-term object segmentation scheme that exploits bidirectional information for change detection is built, based on parametric modelling of global motion. Aspects related to the thresholding procedure, namely the spatial location of foreground candidates and the optimal selection of the involved parameters are examined. Thus, robust segmentation performance is achieved avoiding heuristics and training algorithms for parameter selection. Furthermore, background classification inconsistencies occurring during the independent calculation of segmentation masks over time are addressed using adaptive filtering according to foreground motion. Finally, the exploitation of motion features and object-knowledge on video quality assessment is investigated. Existing objective quality assessment algorithms often rely on the calculation of quality scores ignoring such higher-level information. Thus, possibilities of improving objective video quality assessment models’ performance are herein examined. Specifically, the contributions on objective video quality assessment are threefold; building a content-aware video quality assessment approach that accounts for moving objects, formulating a saliency model that exploits motion features on spatial level and furthermore proposing an approach for consideration of global motion in the temporal dimension that leads to accuracy improvement.

Indexing (document details)
Advisor: Sikora, Thomas
School: Technische Universitaet Berlin (Germany)
School Location: Germany
Source: DAI-C 81/1(E), Dissertation Abstracts International
Subjects: Computer science
Keywords: Computer vision
Publication Number: 10697774
ISBN: 9781392522585
Copyright © 2020 ProQuest LLC. All rights reserved. Terms and Conditions Privacy Policy Cookie Policy