Action recognition with cascaded feature selection and classification

Much of the previous action recognition work focuses on action representation whilst using standard multi-class classifiers such as SVM and k-NN for action classification. We show that these standard classifiers are inadequate in addressing more challenging action recognition problems encountered in an unconstrained environment and propose a novel action classification approach based on cascaded feature selection and classification. Instead of separating multiple action classes simultaneously, the more difficult single task is decomposed automatically into easier sub-tasks of separating two groups of the most separable action classes at a time with different features selected for different sub-tasks. Experiments are carried out using challenging public datasets to demonstrate that with identical action representation, our cascaded classifier significantly outperforms standard multi-class classifiers.