视频的中层视觉表达和高层行为识别研究

项目名称： 视频的中层视觉表达和高层行为识别研究

项目编号： No.61303168

项目类型： 青年科学基金项目

立项/批准年度： 2014

项目学科： 自动化技术、计算机技术

项目作者： 韩志

作者单位： 中国科学院沈阳自动化研究所

项目金额： 28万元

中文摘要： 视觉特征学习、视频表达建模与动作检测识别分别属于计算机视觉研究中底层、中层和高层视觉领域的重要研究课题。目前的研究往往对于不同层面的特定问题提出具有针对性的处理手段，却和其他层面问题脱节，很难统一在一个完整的视频表达体系下。针对该问题，本项目申请旨在以完善视频的中层视觉表达为中心，借鉴格式塔原理等人类视觉系统工作机制，学习完整、有效的底层基元表达字典和部件跟踪方法，在空间和时间两条通道上给出相互关联的中层视觉表达形式与特征。该中层表达不仅能够应用于视频的压缩编码和重建合成，而且通过进一步研究中层表达与高层视觉任务的联系，学习基于中层表达的动作表达模板，并提出高效的动作行为识别算法，从而给出一个全新的、具有高兼容性的多层面视频表达系统。本项目申请的创新性在于从宏观的拟人类视觉系统的角度出发，统一视频处理与理解等问题的理论方法和框架，对于机器视觉系统的统一与集成具有重要的理论意义和应用价值。

中文关键词： 视频基本草图；主动轨迹；本征光照颜色空间；低秩表达；噪声建模

英文摘要： Vision feature learning, video representation modeling and action recognition/detection are the most important issues in the research fields of low-level, middle-level and high-level vision respectively. The previous researches generally solved particular problem in certain level with strong pertinence, however, disconnected with problems in other levels. Therefore, these methods are hardly integrated under a unified video representation framework. Aiming at solving this issue, this project application takes refining middle-level vision representation as the core destination, studies a comprehensive, effective low-level primitive dictionary and component tracking method based on human vision system mechanism, such as Gestalt principle. Given the middle-level vision representation and features in interrelated spatial and temporal tunnels, the middle-level representation can be applied to video coding, compression and synthesis, and with further study on the connection with high-level vision characteristic, it learns a middle-level based action template and proposes an effective action recognition algorithm. Therefore, it gives a brand new, highly compatible multi-level video representation system. The innovation of this project is that, from a more macroscopic aspect of human vision system simulation, it unifies

英文关键词： Video Primal Sketch；Active Trace；Intrinsic Lighting Color Space；Low-rank Representation；Noise Modeling