Towards real-time body pose estimation for presenters in meeting environments

Abstract

This paper describes a computer vision-based approach to body pose estimation. The algorithm can be executed in real-time and processes low resolution, monocular image sequences. A silhouette is extracted and matched against a projection of a 16 DOF human body model. In addition, skin color is used to locate hands and head. No detailed human body model is needed. We evaluate the approach both quantitatively using synthetic image sequences and qualitatively on video test data of short presentations. The algorithm is developed with the aim of using it in the context of a meeting room where the poses of a presenter have to be estimated. The results can be applied in the domain of virtual environments.