Detection of Scene Changes for Video Indexing by Means of the MPEG Motion Vectors

This paper proposes a technique to detect scene changes in a video sequence by means of motion vectors as defined and embedded in the MPEG video codec. Moving objects in a video scene are detected by the MPEG like motion vector search algorithm. Macro-blocks associated with a motion vector are used as image segments that contain moving objects. Feature points defined in terms of salient corners are then determined by the SUSAN principle of corner detection. The K-means clustering is applied to the obtained feature points. Some variants such as the centroid coordinates and average radius are useful indices that indicate the change of video scene. Preliminary results of scene change detection by this proposed method are presented to demonstrate the effectiveness of the method