Human Action Prediction with Hierarchical Movemes

We propose a hierarchical movemes structure for the problem of human action prediction. The features of human actions are calculated by HOG descriptors and exemplar-SVM models. The finer-grained descriptions (movemes) of the human action are calculated from these features using dynamic time warping based segmentation. The dataset we use is collected from Youtube and consists of clips from 20 different TV shows. Different numbers of layers and different classifiers are used to predict the future action. We achieve the best prediction accuracy 46.7% using SVM on the proposed hierarchical moveme model with 3 layers. Keywords— Human Action Prediction, Hierarchical Movemes, Finer Grained Actions, Support Vector Machines, Unsupervised Learning Methods, Dynamic Time Warping segmentation.