Unimodality of MBType frame different metrics for efficient Key Frame Extraction in Video

Abstract

Key Frame Extraction (KFE) is an important block involved with any search process on large scale video logs. KFE has wide applications in fields like Content based retrieval systems, Video Summarization, compression and Video content management. The conventional algorithms exploit the pixel similarities or histogram distributions between frames of video, ignoring the key frame metric information in surveillance videos. The proposed Unimodality of MBType (UMB-KFE) method is motivated by the availability of frame difference metric information of Macro Blocks (I, P, B Frames) and ability to distribute them in a Unimodal distribution form. The difference metrics are further post processed with L2-norm upon the P-frames for efficient reduction. Real merit of the proposed system is that it does not need shot identification, segmentation or context information separation for KFE. Finally, experiments are performed by evaluating the proposed method upon benchmark video datasets and synthetic datasets. A comparison of the results obtained with ground truth information and with state-of-the-art techniques proves that the proposed method is at par in performance by extracting non-repetitive key frames.