摘要
视觉特征提取与特征表达方法在图像分类及识别中十分重要,从特征学习和特征表达角度出发,提出一种基于改进堆叠独立子空间分析模型提取特征的行为识别算法。首先采用两层独立子空间神经网络构建堆叠网络,在特征学习过程中融入正则化约束项,并结合时空卷积算法,获取视频时空层次化不变性特征基元;然后以堆叠卷积网络两层特征基元的非线性映射获取一种规则网格划分下的视频块状局部特征描述符;最后结合时空金字塔匹配模型构建时空层次特征,采用一对多支持向量机分类方法对视频中的动作进行分类。在KTH视频数据库中进行实验。结果表明,该算法学习到的特征基元可对视频构建低维高效的特征描述符,与现有多种行为识别算法进行对比,改进行为识别算法有效性进一步提高。
Aimed at the importance of effective feature extraction and expression method of visual features in an image classification and recognition,a behavior recognition algorithm based on improved stacking independent subspace analysis Model to extract features is proposed. First of all,this algorithm adopts two layers of independent subspace neural networks to form stacked networks. The regularized constrained items are assimilated into in the process of learning features,and the spatio-temporal hierarchical invariant feature primitives of the video are obtained by combining with spatio-temporal convolution algorithm. Then,the video block local feature descriptors are obtained by the nonlinear mapping of the two layer feature primitives of the stacked convolutional network. Finally,the spatio-temporal hierarchical feature descriptors are constructed based on the spatio-temporal pyramid matching model and the actions in the video are classified by using a one to many support vector machine classification method. Experimental results on KTH video database show that the proposed algorithm can form the feature descriptor with low dimension and efficiency,and compared with a variety of existed algorithms,the proposed algorithm is proved to have better effectiveness.
作者
郭晶晶
刘欢欢
GUO Jing-Jing;LIU huan-huan(School of Math and Information,Xinyang University,Xinyang 464000,China)
出处
《软件导刊》
2019年第5期192-196,共5页
Software Guide
基金
信阳学院校级科研项目(2018LYB12)
关键词
行为识别
堆叠独立子空间分析
时空卷积
正则化
时空金字塔
action recognition
stacked independent subspace analysis
spatio-temporal convolution
regularization
spatio-temporal pyramid matching
