Human action recognition by Grassmann manifold learning

In this paper, a kernelized Grassmannian manifold learning method base on new definitions of geodesic distance and manifold graph is proposed to increase inter-class separation and intra-class compactness in human action recognition. Chordal infinite-norm is used to calculate the geodesic distance between subspaces which leads to more inter-class separation. ARMA method is used to describe the spatial-temporal information of the action video. Between-class and within-class similarity graphs are used to map data in a new space. New definition of between class separation graph leads to more separation in the mapped space. The MSR 3D action dataset is used to evaluate the proposed method. The experimental results show robustness of the proposed method.