An Online Object-Based Key Frame Extraction Method for the Abstraction of Surveillance Videos

Abstract

Key frame extraction is an important step for video surveillance. Key frames are able to inform users about the concept of an alarm event and guard environment more efficiently. Key frames can also be used for analysis of feature extraction, indexing and video retrieval. This paper proposes an object-based key frame extraction method for extracting represent- ative frames of an alarm event. The method combines semantic features and weighted importance to extract key frames and devises an object features based formula to obtain better key frames that have clear object image. We also adopt Kalman filter to help predict objects’ situations and extract key frames during the events. The proposed method has been verified by large amounts experiments that include testing 20 clips, implementing in a real-time mobile surveillance system. The experimental videos consist of single objects that are from MPEG-4 test video and so on. Our method proved by experiments not only can get clear and representative key frames but also reduce redundant key frame.