Effective fusing the factored matrices in dual tensors for action recognition

In statistics, Canonical Correlation Analysis (CCA) is a kind of method to effectively analyze the correlation between two data. In recent years, some methods reported in literatures treated the action video sequences as the tensors, and calculated the similarity between two video sequences using CCA. Each of these tensors was unfolded into several matrices or vectors in the previous works. And estimated the similarity between two tensors via accumulating the canonical correlations on all of the pairs of the unfolded matrices or vectors. In this paper, we treat each of the unfolded matrices in a tensor as an individual, instead of accumulating the canonical correlations in a whole tensor, such that we can effectively use the characteristic of each unfolded matrix. We also propose an information fusion method to combine the similarities of each of the unfolded matrices between two tensors. Furthermore, we add the Histogram of Oriented Gradients (HOG) features to complement the tensor generated by the pure video sequence. Our method is validated on the UCF sports database, and the experimental result shows that the proposed method can compete with the state-of-the-art methods.