Obscure Video-Shot Boundary Determination via Protagonist Trajectory Analysis

[[abstract]]Previous research related to Video Indexing had primarily centered on videos with obvious video-shot boundaries, such as movies and recordings of news broadcast. When we encounter a video-shot boundary in the types of film mentioned above, the pixels within the whole scene would alter at a noticeable level, thus allowing the computer to automatically segment and index the video according to the video-shot boundaries. Our research focuses on detecting possible cut points for videos without obvious video-shot boundaries. Although these videos were often recorded continuously without film cuts, the content and people within the film could change rapidly during the recording period. Traditional video-shot boundary detection algorithms cannot successfully detect this kind of content change, which we call as obscure video-shot boundaries, thus we would have to depend on higher level understandings of the film to determine appropriate time points to serve as film cuts. Our proposed algorithm can be roughly classified into two major modules: the moving object segmentation module, and the protagonist-change detection module. Moving object segmentation algorithms can be classified into two categories based on whether the camera moves or not. Our method is applicable only to videos recorded by a static camera, which uses an improved univariate Gaussian background module to perform background subtraction and to extract the foreground moving objects. The largest moving object would be detected and tracked, while its average height and horizontal moving distance would be exploited to determine protagonist change. Whenever a protagonist change occurs, we can conclude that the video content had changed rapidly and meaningfully, thus the time point can be used to serve as a candidate segmentation point. Our proposed algorithm can automate the segmentation and indexing process of those videos without obvious video-shot boundaries, thus greatly reduce the time and effort needed in digital archiving.