Human action recognition based on difference silhouette and static reservoir

Abstract

In this paper, the variation between features of frames for human action recognition is studied, and a new local descriptor extracted among the differences of human silhouettes is posed. This descriptor is represented by coarse histograms based on the distribution of sample points on the outlines of difference silhouettes. The static reservoir is employed as the classifier of human action. Two hyper-parameters, the scaling parameter γ and the regularization parameter C are taken to characterize a static reservoir, and the proper static reservoir for action recognition is identified on the γ−C plane. We test our approach on two commonly used action datasets, and the experimental results show that the proposed method is effective.