zHigh-Order PCA of Video Volume Tensors for Human Action Representation and Recognition.

Abstract: Human action in video sequences can be seen as multidimensional arrays, i.e., tensors, and action video of gray-scale can be directly modeled as a typical 3rd-order tensor. In this paper, we exploit an efficient appearance-based method for recognition of action video volumes using silhouettes based on high order SVD. We use the silhouette ensembles to form a 3rd-order tensor for each action video, search a tensor-based principal component analysis (PCA) to do dimensionality reduction and analyze action volumes. Through experiments on a common database, we see our method not only preserves the spatiotemporal information of the video, but also outperforms the 1-mode PCA in appearance-based recognition of human actions. [Copyright &y& Elsevier]