Multi-layer skeleton fitting for online human motion capture

This paper describes a new approach to fitting a kinematic model to human motion data which is a component of a marker-free optical human motion capture system. Efficient vision-based feature tracking and volume reconstruction by shapefrom-silhouette are applied to raw image data obtained from several synchronized cameras in realtime. The combination of both sources of information enables the application of a new method for fitting a sophisticated multi-layer humanoid skeleton. We present results with real video data that demonstrate that our system runs at 1-2 fps.