Robust Video Content Analysis via Transductive Learning Methods

The amount of video data in the web and in archives§is growing permanently. This fact motivates the§active research area of video indexing, search and§retrieval. Videos vary in many ways: in terms of§recording devices and circumstances, compression§technology, genre, and of course, in terms of§content. The research question addressed by this§Ph.D. thesis is how to build robust video content§analysis approaches that work reliably on arbitrary§videos. For this purpose, a transductive learning§framework is developed that is based on feature§selection and ensemble classification. Apart from§solutions based on the framework, some approaches§employ unsupervised learning or deal with compression§artefacts adequately. Solutions for several video§analysis problems are presented: shot boundary§detection, camera motion estimation, face§recognition, semantic concept detection and semantic§indexing of computer game sequences. Experimental§results on large test sets demonstrate the very good§performance of the approaches. Finally, some areas of§future work are outlined. This thesis is relevant for§students and researchers who are interested in the§field of video content analysis and retrieval.