Pose Estimation of Human Upper Body Using Multi-joint CG Model and Stereo Video Images

We present a body pose estimation method for user interface and telecommunication applications, such as avatar control and desktop operations. We designed our method to meet the following conditions: marker-less, real-time and robust. The estimation is done with twostep matching. First, the acquired image is compared with the prepared CG model images. These model images are pre-generated by using both motion capture data and a camera parameter. When an input image is captured, image matching is done by evaluating the similarity between CG model images and the input images. The joint angles are then estimated by choosing the highest correlated image. Starting with an initial estimation result, joint angle parameters are refined through the iterative model-image synthesis and matching processes. Through this two-step pipeline, we can estimate fine joint angles very robustly in real time. We describe experimental results that show the validity of our approach. The matching process works at 10 frames per second.