Abstract
Video shot transition detection has always been an important and popular research topic because numbers of applications related to video processing, such as key frame extraction, video summarization, require segmenting video into shots as their first step. Basically there are two types of shot transitions: abrupt shot transition (cut) and gradual shot transition including dissolve, wipe and fade. Besides, we also define a special kind of shot transition, called fast-pan, which is mainly caused by fast camera pan action. In the thesis, we proposed a learning-based shot transition detection system to accomplish this work. For cut detection subsystem, color-based and motion-based features are extracted. In the gradual transition detection subsystem, the luminance-based and edge-based features are added since it is more complicated than cut. Motion-based and gradient-based features are employed in fast-pan detection subsystem. By separately applying these features into a learning machine, we can train three different classifiers to detect cuts, gradual transitions and fast-pan events individually. Finally the experimental results are shown. Our experimental results give excellent detection accuracy for all the three shot transition subsystems. The performance of the proposed system on the TRECVID 2003 benchmarking videos for cut and gradual shot transition detection is comparable to the best results reported in the competition.