Video scene segmentation using the shot transition detection by local characterization of the points of interest

In this article, we propose a new approach to scene segmentation of video based on the shot detection. The methods of segmentation by scenes are still expensive in computation, in most of time; this detection uses matrices of dissimilarity as calculation kernel for clustering algorithms, which implies a quadratic nature of these calculations. However, shot segmentation algorithms are generally linear complexity. To take advantage of the benefits of both methods, our idea is to apply two-level mixing of these kinds of segmentation algorithms. For this, we define a measure of dissimilarity characterized by points of interest determined from the frames of the video. The signature based on these points allows us to consider the sequence as a time series, which we try to model the behavior by a measure of dissimilarity. Our segmentation method is to cut the video by shot, to set the intervals of the shots and then retrieve the representative images (or key frames) of these shots. Last time we apply an automatized clustering algorithm on this set of keyframes to join the plans represented by similar key frames. The latter operation will group the shots of similar contents, but distributed in the video, on other words, the joint of the same contents shots constitutes a scene. Some experimental results on real sequences show the validity of our approach.