In this paper

a simple but efficient algorithm based on tensor subspace learning is proposed to reduce the dimensionality of high-dimensional silhouette data for human action recognition. For image sequences of each action

they are projected into a low dimensional subspace so that both spatial and temporal properties of the action are preserved. Further

a nearest-neighbor action recognition is carried out basing on Hausdorff distance. Two experiments for action recognition and robust test have been carried out to testify the effectiveness of introduced tensor subspace learning.