人体动作识别中基于HTM架构的时空特征提取方法

针对人体动作识别中时空特征提取问题,提出一种基于层次时间记忆(HTM)架构的深度学习模型,用来提取图像帧的时空特征。首先,将图像帧构建成树形节点层次结构。然后,在每一层中,通过欧氏距离分组来提取图像样本的空间特征,利用时间邻接矩阵提取样本的时间特征,利用置信传播方法将各层局部特征组进行汇总归类,最终得到整体特征组,作为该图像帧的时空特征。此外,在节点操作中引入张量代数,从而避免出现高维特征。最后,将特征送入支持向量机(SVM)分类器进行识别分类。在MSR Gesture 3D动作数据库和KTH动作数据库上的实验结果表明,提出的方法能够有效提取出高分类性能的时空特征,分类准确率高于其他几种较新的方法。

Spatial-Temporal Feature Extraction Method Based on HTM Structure for Human Action Recognition

For the issue that the spatial-temporal feature extraction for human motion recognition, this paper proposes a depth learning model based on the hierarchical time memory (HTM) architecture to extract the spatial-temporal features of the image frame. Firstly, the image frame is constructed as the tree hierarchy structure. Then, it uses of clustering based on Euclidean distance to extract the spatial features of sample image, and uses of time adjacency matrix to extract the temporal features. At the same time, it adopts belief propagation method to gather the local feature group, so as to obtain the overall feature as the spatial-temporal feature of the image frame. In addition, the tensor algebra is integrated into the node operation to avoid the occurrence of high dimensional features. Finally, the feature is inputted into the support vector machine (SVM) classifier for recognition and classification. The experimental results on the two database MSR Gesture 3D action database and KTH action database show that proposed method can effectively extract the spatial-temporal feature with high classification performance, and it has higher classification accuracy rate than several other advanced methods.