基于深度学习和事件驱动的视频聚合研究

Video analysis has received significant attention from academic and industry both in China and abroad. Recently, video mashup becomes one of the hottest research topic. The widespread of wearable cameras makes it possible to record egocentric videos. Also, the explosive growth of video content on the Web has been revolutionizing the way people share, exchange and perceive videos. Videos are uploaded by different users at different locations and/or times can embody different emphasis and compensate each other in describing the event. Combining these videos from different sources together can unveil a more comprehensive picture of the event. However, simply combining the videos may degrade user experience in that it is time-consuming and tedious to view those highly redundant, noisy and disorganized content. Thus, it is necessary to cut and re-join the videos to render users the big picture of the events in a more logical and compact way. Most of the existing video mashup methods are focusing on structured videos or a single video. In this project, we will focus on studying and proposing a set of deep learning and event driven video mashup methods to improve the accuracy, efficiency and reliability of video mashup. Specifically, this project consists three stages: 1) integrating deep learning technique and optimized video cut algorithms to facilitate egocentric video based key shot selection; 2) proposing event driven video mashup; and 3) automatically generating a single unified mashup video from a collection of relevant video and images. The goal of this project is to propose a novel video mashup framework, which will contribute to real world video dataset management applications.

基于视频分析的研究一直受到国内外学术界和工业界的极大重视。近年来，视频聚合成为了该领域的一个研究热点并被广泛研究,但已有的方法主要集中于分析有结构的视频。与传统的结构化的视频相比，以自我为中心的视频及网络视频为代表的视频具有无结构性和海量性。这些特性导致目前已有的视频聚合方法无法直接应用于该类视频的聚合。深度学习作为一种新的方法被广泛的运用于机器学习和多媒体领域，具有非常出色的特征提取和分类的能力。目前，事件驱动的视频聚合为海量网络视频的聚合提供了一个很好的切入点。本项目从三个方面研究基于深度学习和事件驱动的视频聚合方法：一、基于深度学习的以自我为中心的视频关键片段选取算法及优化；二、深入分析基于事件驱动的网络视频聚合方法；三、基于跨媒体的事件驱动的视频聚合算法及优化。本项目的研究切合当下视频研究的趋势，对视频聚合这一领域具有重要的意义，为实现高效的视频管理与视频信息抽取奠定了基础。

基于视频分析的研究一直受到国内外学术界和工业界的极大重视。近年来，视频聚合成为了该领域的一个研究热点并被广泛研究,但已有的方法主要集中于分析有结构的视频。与传统的结构化的视频相比，以自我为中心的视频及网络视频为代表的视频具有无结构性和海量性。这些特性导致目前已有的视频聚合方法无法直接应用于该类视频的聚合。深度学习作为一种新的方法被广泛的运用于机器学习和多媒体领域，具有非常出色的特征提取和分类的能力。目前，事件驱动的视频聚合为海量网络视频的聚合提供了一个很好的切入点。本项目从三个方面研究基于深度学习和事件驱动的视频聚合方法：一、基于深度学习的以自我为中心的视频关键片段选取算法及优化；二、深入分析基于事件驱动的网络视频聚合方法；三、基于跨媒体的事件驱动的视频聚合算法及优化。本项目的研究切合当下视频研究的趋势，对视频聚合这一领域具有重要的意义，为实现高效的视频管理与视频信息抽取奠定了基础。