OSiMa: Human pose estimation from a single image

Abstract

Human upper body pose estimation plays a key role in applications related to human-computer interactions. We propose to develop an avatar based video conferencing system where a user's avatar is animated following his/her gestures. Tracking gestures calls for human pose estimation through image based measurements. Our work is motivated by the pictorial structures approach and we use a 2D model as a collection of rectangular body parts. Stochastic search iterations are used to estimate the angles between these body parts through Orientation Similarity Maximization (OSiMa) along the outline of the body model. The proposed approach is validated on human upper body images with varying levels of background clutter and has shown (near) accurate pose estimation results in real time. © 2011 Springer-Verlag Berlin Heidelberg.