A compact discriminant hierarchical clustering approach for action recognition

In order to improve the accuracy of action recognition, a compact discriminant hierarchical clustering approach and an action recognition new framework are respectively proposed. Firstly, on the bases of low-level features 3D Self-Correlation Histogram of Oriented Gradient in Trajectory (3D_SCHOGT) and 3D Self-Correlation Histogram of Oriented Optical Flow in Trajectory (3D_SCHOOFT), the mid-level semantics possessing purity, representativeness and discriminativeness simultaneously are obtained using the proposed compact discriminant hierarchical clustering approach, in which removal of singularities, quantitative evaluations of purity, representativeness and discriminativeness, as well as additive constraint of information entropies for clusters are conducted respectively to assure the better purity, representativeness and discriminativeness. Secondly, by introducing category constraint, a discriminant classification model of Category Constraint Latent Support Vector Machines (CC-LSVM) is proposed, which enhances the discriminative ability of classifier. Finally, to further improve the accuracy of action recognition, a new framework is proposed, which introduces low-level features, mid-level semantics and mid-level semantic self-correlation features into the proposed CC-LSVM classifier in a weighted association way, makes full use of category information of actions, and mines the correlations between multi-semantic features and action categories. Consequently, the action recognition accuracy is improved. The accuracies on Weizmann, KTH, UCF-Sports and YouTube datasets are 100%, 98.83%, 98.67% and 90.73% respectively, which outperform all those in contrastive methods. Experiments demonstrate the effectiveness of proposed compact discriminant hierarchical clustering approach and new framework.