direkt zum Inhalt springen

direkt zum Hauptnavigationsmenü

Sie sind hier

TU Berlin

Inhalt des Dokuments


Towards Robust Object Segmentation in Video Sequences and its Applications
Zitatschlüssel 1292Karaman2010
Autor Mustafa Karaman
Jahr 2010
Notiz Nr. 25
Schule Technische Universität Berlin
Zusammenfassung The primary goal of this thesis is to develop a robust object segmentation method for videos acquired from static cameras typically used in surveillance applications. Therefore, it explores and improves background modeling and subtraction techniques. The first part of the dissertation deals with the pre-processing stage of the general object segmentation framework in which a new hierarchical background estimation method is proposed to create a reliable background model even in the presence of moving objects. For the main-processing stage the thesis explores and compares nine state-of-the-art methods using subjective and objective evaluation measures. The advantages and disadvantages of the different methods are analyzed and discussed. Based on that, the most promising ideas are adopted for a more reliable background subtraction method. It combines the invariant color description of a Gaussian color model (GCM) with a temporal consistency criterion. Since reflections of an object on the ground are a major challenge for any object segmentation approach, a reflection detection and removal method is proposed for the post-processing stage. It analyzes the appearance of the over-segmented object to detect the reflection boundary and adapt the object mask accordingly. Furtheremore a new framework for figure-ground image segmentation evaluation is proposed to compare different image segmentation approaches automatically. The developed object segmentation framework has been used in several applications to prove its versatility. Finally, a personalized human computer interface, that detects, tracks and identifies persons and recognizes their gestures being used in an intelligent cash machine scenario, has been developed based on the combination of visual appearance and motion based analysis.
Download Bibtex Eintrag

Zusatzinformationen / Extras


Schnellnavigation zur Seite über Nummerneingabe