Within this paper a technique for model-based 3D hand tracking is presented. A hand model is built from a set of truncated quadrics, approximating the anatomy of a real hand with few parameters. Given that the projection of a quadric onto the image plane is a conic, the contours can be generated efficiently. These model contours are used as shape templates to evaluate possible matches in the current frame. The evaluation is done within a hierarchical Bayesian filtering framework, where the posterior distribution is computed efficiently using a tree of templates. We demonstrate the effectiveness of the technique by using it for tracking 3D articulated and non-rigid hand motion from monocular video sequences in front of a cluttered background.