Tensor-based projection using ridge regression and its application to action classification

Abstract

In this study, the authors encode video sequence by a third-order tensor instead of matrices or vectors and present a novel framework of action classification via tensor-based projection using ridge regression (TPRR). The main contributions of this framework include: (i) Classification can be directly achieved by projecting the tensor samples into their binary class label through ridge regression technique. (ii) TPRR algorithm is further generalised from binary classification to solve multi-category classification problem by incorporating the error correcting concept. Experiments of action classification on a small or medium training set demonstrate that this approach outperforms previous tensor embedding methods and other state-of-the-art techniques, whereas on large training set, it is also able to achieve very satisfactory results. The proposed approach is further proven to be considerably robust to viewpoint, partial occlusion and irregularities in motion styles.