多尺度特征计算加速算法与同类行为视角无关描述符挖掘方法研究

项目名称： 多尺度特征计算加速算法与同类行为视角无关描述符挖掘方法研究

项目编号： No.61472196

项目类型： 面上项目

立项/批准年度： 2015

项目学科： 计算机科学学科

项目作者： 刘云

作者单位： 青岛科技大学

项目金额： 81万元

中文摘要： 不同视图的同类行为对象具有尺度变化和视角各异的鲜明区别，是影响行为识别的关键因素。视频中的多尺度特征计算，通常采用精细的尺度金字塔分层策略分别计算每层的底层特征。但该方案提取的信息冗余而实时性不足，已成为实际应用亟待解决的瓶颈。本课题首先提出基于特征预测算法的金字塔分层策略，加速多尺度特征的计算，从根本上解决实时性不足问题。其次针对同类行为不同视角带来的行为识别精度问题，拟采用非线性动力学系统的递归图分析方法，通过建立与系统原相空间等价的相空间递归图，根据递归图对角线方向具有发育较好的混动递归线条纹理的属性，挖掘其递归相似性。然后针对传统聚类方法中须预置类别个数以及不能有效聚类非凸集数据集合等缺点，拟采用基于流形相似度计算的多智能体进化聚类算法，对挖掘出的行为递归相似性数据集进行结构性分析。最后，基于关键词词袋策略建立同类行为模型。本研究为计算机视觉中行为分析和识别的实际应用提供理论依据。

中文关键词： 多尺度特征计算实时性；多视图视频相关性；递归图分析；多智能体进化；行为识别

英文摘要： View-dependent videos for one same action are quite visually different at multi-sacle changes and view-point variations, which are crucial factors for action recognitions. Calculations of multiscale features from video images are usually based on finer scale pyramid strategy, which is to figure out low-level features respectively at each scale leavel. This scheme is of redundent information but worse of reatime, which becomes urgent bottleneck in applications. A new fast scale pyramid strategy based on feature forecast algorithm is proposed, which can speed up low-level feature calculations and solve real-time problem fundamentally. Secondly to overcome action recognition inaccuracy problem caused by view-point variations, Recurrence Plot method primarily for none linear dynamics system is adopted to construct phase space image, which is equivalent as system original phase space and can be exploited to mine recurrent similarities from view dependent videos alone its diagonal, where chaotic recurrent texture features are well developed. Traditional clustering method cannot solve none convex-distributed data sets, further more its number of clustering types have to be pre-decided that will destroy data original structure, a muti-agent evolutionary clustering algorithm is proposed which is sensitive to data distribution shape, so that the true action pattern structure can be explicitly discovered from the above various view-dependent video recurrent features. Finally one action model can be established based on bag of words. This study provides a theoretical basis for the practical application of action analysis and recognition in computer vision.

英文关键词： Real Time Calculation of Multiscale Features;Correlations of Multi-view Action Videos;Recurrence Plot Analysis;Multi-agent Evolution;Action Recognition