Algorithms for Video Structuring

Video structuring aims at automatically finding structure in a video sequence. Occupying a key-position within video analysis, it is a fundamental step for quality indexing and browsing. As a low level video analysis, video structuring can be seen as a serial process which includes (i) shot boundary detection, (ii) video shot feature extraction and (iii) video shot clustering. The resulting analysis serves as the base for higher level processing such as content-based image retrieval or semantic indexing. In this study, the whole process is examined and implemented. Two shot boundary detectors based on motion estimation and color distribution analysis are designed. Based on recent advances in machine learning, a novel technique for video shot clustering is presented. Typical approaches for segmenting and clustering shots use graph analysis, with split and merge algorithms for finding subgraphs corresponding to different scenes. In this work, the clustering algorithm is based on a spectral method which has proven its efficiency in still-image segmentation. This technique clusters points (in our case features extracted from video shots) using eigenvectors of matrices derived from data. Relevant data depends of the quality of feature extraction. After stating the main problems of video structuring, solutions are proposed defining an heuristical distance metric for similarity between shots. We combine color visual features with time constraints. The entire process of video structuring is tested on a ten hours home video database.