Dynamic selection and effective compression of key frames for video abstraction

This paper reports on a new key frame based video abstraction method. With our method, a video sequence is first segmented into a number of video shots. Several key frames are selected in each shot using a dynamic selection technique. For these key frames, a motion-based clustering algorithm is applied so that key frames in the same cluster are alike in sense of motion compensation error, while those from different clusters are quit dissimilar. Then a novel cluster-based coding scheme is developed for efficient representation of the key frames. Simulations show that the proposed method can select key frames according to the dynamics of a video sequence and abstract the video with different levels of scalability.