基于张量学习的多源异质多视角视频显著性分析

项目名称： 基于张量学习的多源异质多视角视频显著性分析

项目编号： No.61503235

项目类型： 青年科学基金项目

立项/批准年度： 2016

项目学科： 自动化技术、计算机技术

项目作者： 朱国康

作者单位： 安徽大学

项目金额： 22万元

中文摘要： 视频显著性分析旨在自动提取视频中用户可能感兴趣或最具判别价值的信息，以便后续更有针对性地进行重点处理与分析，是建立智能安全防范系统的重要途径之一。其中，对视频中目标的特征描述、视觉反差度量是显著性分析的两个核心内容。随着技术的进步与需求的提升，视频监控数据的多源异质多视角特性日益凸显，而现有方法缺乏对多源异质多视角视频的联合建模与关联挖掘。本项目着眼于信息的关联表达与协同作用，在张量学习的理论框架下研究显著性分析，包括：1）基于高阶张量约束，进行视频间时空关系的准确估计，据此建立统一参考系；2）基于张量学习理论，从视频数据中挖掘和表达目标的空间特征；3）基于张量流形上的哈希索引，快速提取运动信息；4）基于张量原型空间，构造多种视觉敏感特征的反差度量。本项目的研究不仅可以拓展现有显著性分析的研究范式，丰富张量学习理论，同时也将进一步促进视频监控系统的智能化以更好地服务于社会公共安全事业。

中文关键词： 生物视觉模型；注视机理；视觉建模

英文摘要： Reliable salient information analysis can entail the video processing and analysis techniques the selection ability to optimize resource allocation for better efficiency, and therefore can contribute importantly to intelligent security systems. There are two key components in saliency detection, including feature description and measure of visual contrast. With the development of science and technology, the videos that people can acquire have been expanded to heterogeneity and multi-view. However, the existing saliency detection methods are powerless for joint modeling and association mining of the expanded information. This proposal exploits information from multiple heterogeneous and multi-view videos jointly on the basis of tensor learning. The main innovative ideas include: 1) constructing a common reference system for the multiple heterogeneous and multi-view videos by high-order tensor constraints, through which the temporal-spatial relationships between couples of videos could be estimated more accurately; 2) learning a informative yet compact feature of static saliency from multiple heterogeneous and multi-view videos via tensor learning; 3) developing a hashing model on tensor manifold for efficient motion information extraction; 4) establishing a tensor prototype space for the measure of tensor-based visual feature contrast. The research of this proposal has the potential to enrich the paradigms of saliency analysis, and expand tensor learning theory. Besides, it can also benefit intelligent security systems for better public safety.

英文关键词： bio-inspired vision model;fixation mechanism;vision modeling