A Deterministic Key-Frame Indexing and Selection for Surveillance Video Summarization

Abstract

Video data is voluminous and impacts the data storage devices as there are CCTV surveillance videos being created every minute and stored continuously. Due to this increase in data there is a need to create semantic information out of the frames that are being stored. Video Summarization is a process that continuously monitors changes and helps in reducing the number of frames being stored. This work enables summarization to be carried out based on selecting threshold-based system that can select key-frames ideally suit for storage and further analysis. Initially a Global threshold based on Otsus method is carried out for all frames of a surveillance video and based on the set threshold a retrospective comparison is done on each frame based on statistical methods to converge on determining the keyframes. A similarity index is generated based on the iterative comparison of frames based on global and local threshold comparison. The local threshold is indexed based on Analysing Method Patterns to Locate Errors(AMPLE), An-derbergs D(AbD), Cohens Kappa(CK), Tanimoto Similarity(TS), Tversky feature contrast model(TFCM), Pearson coefficient of mean square contingency(Pmsc). The Global threshold is updated each time a keyframe is selected based on the comparison of local and global threshold. The results are compared with five surveillance videos and six methods to identify keyframes Selection Rate is the metric used for calculating the performance.