F-DES: Fast and Deep Event Summarization

Abstract

In the multimedia era a large volume of video data can be recorded during a certain period of time by multiple cameras. Such a rapid growth of video data requires both effective and efficient multiview video summarization techniques. The users can quickly browse and comprehend a large amount of audiovisual data. It is very difficult in real-time to manage and access the huge amount of video-content-handling issues of interview dependencies significant variations in illumination and presence of many unimportant frames with low activity. In this paper we propose a local-alignment-based FASTA approach to summarize the events in multiview videos as a solution of the aforementioned problems. A deep learning framework is used to extract the features to resolve the problem of variations in illumination and to remove fine texture details and detect the objects in a frame. Interview dependencies among multiple views of video are then captured via the FASTA algorithm through local alignment. Finally object tracking is applied to extract the frames with low activity. Subjective as well as objective evaluations clearly indicate the effectiveness of the proposed approach. Experiments show that the proposed summarization method successfully reduces the video content while keeping momentous information in the form of events. A computing analysis of the system also shows that it meets the requirement of real-time applications.