基于深度序列的时空金字塔的动作识别

摘要

An efficient human action recognition method is put forward. Three views of depth sequences are transformed into Depth Motion Outline Sequence(DMOS)by using the method of interframe differentiation. Then a spatio-temporal pyramid is proposed to subdivide the DMOS on temporal and spatial level. A feature fusion scheme is presented to concat-enate the Histograms of Oriented Gradients(HOG)features which have extracted from the subdivided DMOS. Finally lin-ear SVM to classification is used. Through using the MSR Action 3D data sets, this method is evaluated with different parameters of spatial-temporal pyramid. Experimental results show that this method has higher recognition rate than the similar algorithm.%提出一种高效的人体动作识别方法.通过帧间差分法将深度序列的三视图转化为深度运动轮廓序列(DMOS),然后利用时空金字塔对DMOS进行时间维和空间维细分,将细分后得到的空间网格的局部方向梯度直方图(HOG)进行特征融合,并使用线性SVM分类.最后采用MSR Action 3D数据集对提出的算法在不同时空金字塔参数下的识别率和处理速度进行了评估,结果表明该方法在同类算法中具有更高的识别率.