EVS-DK: Event video skimming using deep keyframe

摘要

In this automation era, video surveillance becomes an essential component and omnipresent at ATMs, public places, airports, railways, roadways, etc. There are many challenges to store and access such massive data generated by video surveillance. Therefore, a novel technique is required to manage the comprehensive view of the content. In this work, we propose an event summarization technique using Deep learning framework for monocular videos. A spatiotemporal similarity function is developed to construct a similarity matrix based on the visual features. Video frames are represented by the sparse matrix as graph vertices based on an objective function, where Highly Connected Subgraphs (HCS) are constructed as clusters. Finally, events are obtained from such clusters assuming that the centroid of the cluster is a key-frame of the event. Consequently, this approach does not require assumption to determine the number of clusters. Due to this advantage, users can select the number of keyframes without incurring an extra computational cost. Experimental results on two benchmark datasets show that the proposed model outperforms the state-of-the-art models on Precision and F-measure and also cover the major contents of the original video. (C) 2018 Elsevier Inc, All rights reserved.