Abstract

Pose-based features have demonstrated to outperform low-levelappearance features in human action recognition. New RGB-D cameras provide locations of human joints with which geometric correspondences can be easily calculated. In this article, a new geometric correspondence between joints called Trisarea feature is presented. It is defined as the area of the triangle formed by three joints. Relevant triangles describing human pose are identified and it is shown how the variation over time of the selected Trisarea features constitutes a descriptor of human action. Experimental results show a comparison with other methods and demonstrate how this Trisarea-based representation can be applied to human action recognition.