On-the-fly Extraction of Key Frames for Efficient Video Summarization

We propose in this paper an object-based method for on-the-fly extraction of key frames summarizing the salient visual content of videos. This method is based on spatial segmentation of each frame in order to detect the important events. Thus, key frame detection is facing a much more semantic criterion so that each key frame presents an important event such as the appearance and the disappearance of relevant objects. Realized experiments on challenging videos demonstrate the efficiency of the proposed method, which is able to catch the semantic content of a video shot, while preventing redundancy of the extracted key frames and maintaining minimum requirements in terms of memory space.