Shot Boundary Detection from Video Sequence Using Motion Direction Histogram Features

In this paper, we propose "the motion direction histogram features" as new features for the shot boundary detection from video sequence. Also we demonstrate the effectiveness of our method through the experimental results. The shot boundary detection from video sequence is an important task for video handling technology. Because dividing the video into the shot is first step of video handling. However, the conventional method had the problem that mistook the shot with the intense whole screen movement for the shot boundary, for example the shot that tracks the large object that moves at high speed and the shot that has the high-speed panning. The cause of the problem is the conventional features taking large value when the shot boundary detector faces such motion. To overcome this problem, we propose "the motion direction histogram features." The features express the difference between directional distributions of the motion vectors by the absolute value difference sum of the histogram. The effectivenss of the features was originated from difference of directional distribution of the motion vectors between shot boundaries and such motion. As a result of the shot boundary detection experiment, the detection accuracy improved 26% at the maximum, by integrating the motion direction histogram features and conventional features using the discrimination function. And, we found that the distribution of features differed with each genre when we analyzed this experimental result. In addition, the detection accuracy improved 6% at the maximum by making the discrimination dictionary in each genre.