Improving the Bag-of-words Model Using the Hubness Phenomenon for Skeleton-based Action Recognition

Abstract

This paper proposes a new bag-of-words (BoW) approach for human skeletal action recognition based on the hubness phenomenon. Many previous action recognition methods adopt a BoW method that ignores the high dimensions of the feature space. Hence, this paper presents a visual vocabulary construction approach for action recognition. The action feature descriptors including the spatial-temporal joint information are first extracted from action sequences. Various factors are considered in the spatial descriptors, including the three-dimensional coordinates of the joints, angles between body segments, and direction, as well as elevation angles. The feature vectors with higher hubness scores are then selected as candidates. Finally, the visual words are selected from the candidates using the maximin method. We evaluated the effectiveness of the proposed method on two public datasets, the CAD-60 and UTKinect datasets. The experimental results show that the proposed method performs better than several state-of-the-art algorithms.