A Feature Discriminability Focusing Model for Skeleton-based Human Action Recognition.

In order to deal with the complexity of human movements and the variety among the same action performed by distinct subjects, previous approaches for human recognition usually use multi-type features to get more powerful discriminability for improved performance on the condition that the different types of features have the equal strength for each class. But in the real situations, the fact may be that different types of features and/or some dimensions of an individual feature could have distinct discriminability or contribution while classifying different categories of human actions. In this work, we propose a novel Feature Discriminability Focusing Model (FDFM) to focus on the features and/or dimensions with stronger discriminability for integrating multiple features. Specifically, a weight matrix derived from sparsity-inducing norms is leant for every feature with respect to each category individually, in which each element is regarded as a saliency indicator to suggest the contribution of using the feature to make a correct classification. A new optimized algorithm is designed to solve the proposed model. By coupling with three types of simple skeletal features, our proposed model achieves competitive performance on three widely used benchmarks.