An efficient algorithm for scene detection is proposed. Where firstly within a sliding shot window

the shots are clustered and each cluster center is achieved rapidly by employing MS clustering

then according to the development pattern of film video scene

the temporal distances between two shots are computed and scenes are detected based on the temporal and spatial relationship. In addition

the scene key frames can be achieved on the basis of corresponding cluster centers. Finally

a succeeding procession for over-segmented scenes is introduced. Experiments prove this algorithm can cluster shots rapidly and detect scenes and the key frames of the scence efficiently.