海量网络视频中的复杂事件检测技术研究

项目名称： 海量网络视频中的复杂事件检测技术研究

项目编号： No.61201387

项目类型： 青年科学基金项目

立项/批准年度： 2013

项目学科： 电子学与信息系统

项目作者： 姜育刚

作者单位： 复旦大学

项目金额： 26万元

中文摘要： 本项目针对当前海量网络视频内容分析这一重大需求，提出一套完整的视频事件检测技术，包括视频事件训练数据的自动采集、多模态视频特征表示和基于上下文分析的事件学习算法。在训练数据采集方面，拟提出一种基于多重文本相似度的度量方法，对网络关键词检索结果进行过滤，进而得到高精度的事件标注；在多模态视频特征表示方面，采用图聚类方式生成视音频联合词袋，深入探索模态间的关联关系；此外，本项目将设计事件检测的上下文分析算法，利用基本概念（如目标、场景）的检测结果提高复杂事件的检测精度。该算法采用有向图来对事件-概念关系建模，以充分发掘事件-概念间的因果及共生关系。本项目的研究成果将为网络视频内容分析奠定一定的理论基础，并为网络视频检索、内容监管等一系列重要应用提供系统化解决思路。研究的成果也将通过国际权威视频分析评测活动检验其性能（如美国国家标准局的视频检索评测TRECVID）。

中文关键词： 视频内容识别；数据采集；特征提取；；

英文摘要： This project will focus on an emerging problem in the area of large-scale Internet video analysis. A set of algorithms will be devised for detecting complex Internet video events, including automatic training set collection, multimodality video feature representation, and context-based event learning algorithms. In training set collection, we plan to design a method that uses multiple textual similarity measurements to evaluate the relevancy between an evant and a candidate video. This will be used to filter the noisy web search results and generate a clean set of training samples. In multimodality feature representation, we will use graph-based clustering methods to produce audio-visual joint codebooks, which, compared with using the two modalities separately, can explore deep correlations between them and therefore may be more effective for video event detection. Finally, for the context-based event learning, we plan to model event-concept relationships using a directed graph. Here the concepts refer to lower-level semantic entities such as basic objects and scenes, which are useful cues for event detection in videos. The event-concept relationships will be used in a graph-based formulation that refines initial event detection results, using a formulation that is efficient to solve. The research outcomes of th

英文关键词： Video Content Recognition；Data Collection；Feature Extraction；；