Video object segmentation.

详细信息   

• 作者：Porikli ; Fatih Murat.
• 学历：Doctor
• 年：2002
• 导师：Wang, Yao
• 毕业院校：Polytechnic University
• 专业：Engineering, Electronics and Electrical.;Computer Science.
• ISBN：0493590900
• CBH：3045267
• Country：USA
• 语种：English
• FileSize：7680836
• Pages：218

文摘

In this thesis, the problems associated with the automatic object segmentation of the video sequences are considered. Towards this goal, a unique framework that combines various disciplines of image and video processing techniques ranged from noise filtering to data clustering is developed. The framework also addresses a number of challenging issues associated with computational complexity, accuracy, generality, and robustness.;One of the primary aims of this thesis is to study the fusion of color, texture, motion, shape, frame difference, and other methods of video segmentation for automatic detection considering the real-time processing requirements. In contrast to frame-wise tracking techniques, the employment of a spatiotemporal data that is constructed from multiple video frames introduces new degrees of freedom that can be exploited in terms of object extraction and content analysis. The current notions of region segmentation are extended to the spatiotemporal domain, and new models to estimate the object motion are derived.;Another objective of this work is to explore techniques and algorithms that provide efficient means of preprocessing of input video sequence. The newly designed image simplification and reconstruction filters enable us to develop efficient algorithms for noise removal, and they prevent from over segmentation difficulties. The problem of adaptation of system parameters and thresholds is formulated and solved. Also, the relation between the color space and the similarity functions is investigated.;As a final objective of this thesis, a clustering problem that considers construction of meaningful video objects from color homogeneous regions is examined. Specifically, the fine-to-coarse and coarse-to-fine strategies are discussed. Novel de scriptors to evaluate the quantitative and relational attributes of the extracted objects are introduced. Additionally, new sources of motion information is considered, such as the trajectory definition of an object. Also, the area of multi-resolution object representation is explored.