文章基本信息

标题：A Novel Topic Extraction Method based on Bursts in Videos Streams
本地全文：下载
作者：Kimiaki Shirahama ; Kuniaki Uehara
期刊名称：International Journal of Hybrid Information Technology
印刷版ISSN：1738-9968
出版年度：2008
卷号：1
期号：3
出版社：SERSC
摘要：In this paper, we introduce a novel method for extracting "topics" as interesting events in a video. Here, we define the interestingness of an event by the anomaly of a target character's appearance and disappearance pattern. As examples of abnormal patterns, shot durations in thrilling events are very short while shot durations in romantic events are very long. In contrast, as an example of non-abnormal pattern, conversation events are presented by the pattern, where the target character repeatedly appears in one shot and then another character appears in the next shot. From the above point of view, our topic extraction method aims to detect the following two types of abnormal patterns, called "bursts". The first type of burst is a pattern where the target character appears in shots with very short durations, while the second is a pattern where he/she appears in shots with very long durations. To detect such bursts, we firstly divide the video into events characterized by specific patterns of the target character's appearance and disappearance. We locate these patterns in the video by using time series segmentation technique. Then, we extract topics by examining whether the pattern in each event can be regarded as a burst or not. Experiments on different videos validate that a character's appearance and disappearance patterns are effective for obtaining semantically meaningful events. And, bursts are useful for extracting many interesting topics.