【摘要】 为充分利用时空分布信息及视觉单词间的关联信息,提出了一种新的时空非负成分表示方法(ST-NCR)用于动作识别.首先,基于视觉词袋(Bo VW)表示,利用混合高斯模型对每个视觉单词所包含的局部特征的时空位置分布进行建模,计算时空Fisher向量(STFV)来描述特征位置的时空分布;然后,利用非负矩阵分解从Bo VW表示中学习动作基元并对动作视频进行编码.为有效融合时空信息,采用基于图正则化的非负矩阵分解,并且将STFV作为图正则化项的一部分.在3个公共数据库上对该方法进行了测试,结果表明,相比于Bo VW表示和不带时空信息的非负成分表示方法,该方法能够提高动作识别率.更多还原

【Abstract】 To make full use of spatial-temporal information and the relationship among different visual words,a novel spatial-temporal nonnegative component representation method( ST-NCR) is proposed for action recognition. First,based on Bo VW( bag of visual words) representation,the locations of local features belonging to each visual word are modeled with the Gaussian mixture model,and a spatio-temporal Fisher vector( STFV) is calculated to describe the location distribution of local features. Then,nonnegative matrix factorization( NMF) is employed to learn the action components and encode the action video samples. To incorporate the spatial-temporal cues for final representation,the graph regularized NMF( GNMF) is adopted,and STFV is used as part of graph regularization. The proposed method is extensively evaluated on three public datasets. Experimental results demonstrate that compared with Bo VW representation and nonnegative component representation without spatio-temporal information,the method can obtain better action recognition accuracy.更多还原