A spatial-temporal iterative tensor decomposition technique for action and gesture recognition

Classification of video sequences is an important task with many applications in video search and action recognition. As opposed to some traditional approaches that transform original video sequences into forms of visual feature vectors, tensor-based methods have been proposed for classifying video sequences with natural representation of original data. However, one obvious limitation of tensor-based methods is that the input video sequences are often required to be preprocessed with a unified length of time. In this paper, we propose a technique for handling classification of video sequences in unequal length of time, namely Spatial-Temporal Iterative Tensor Decomposition (S-TITD) for uniform length. The proposed framework contains two primary steps. We first represent original video sequences as a third-order tensor and perform Tucker-2 decomposition to obtain the reduced-dimension core tensor. Then we encode the third order of core tensor to a uniform length by adaptively selecting the most informative slices. Notably, the above two steps are embedded into a dynamic learning framework to guarantee the proposed method has the ability of updating results over time. We conduct a series of experiments on three public datasets in gesture and action recognition, and the experimental results show that the proposed S-TITD approach achieves better performances than the state-of-the-art algorithms.