Viewpoint-Aware Action Recognition Using Skeleton-Based Features from Still Images

In this paper, we propose a viewpoint-aware action recognition method using skeleton-based features from static images. Our consists of three main steps. First, categorize the viewpoint an input image. Second, extract 2D/3D joints state-of-the-art convolutional neural networks and analyze geometric relationships for computing 2D 3D skeleton features. Finally, perform view-specific classification per person, based on categorization extracted We implement two multi-view data acquisition systems create new dataset containing labels, in order to train validate our method. The robustness proposed changes was quantitatively confirmed datasets. A real-world application recognizing various actions also qualitatively demonstrated.