3D hand tracking in a Stochastic approximation setting

This paper introduces a hand tracking system with a theoretical proof of convergence. The tracking system follows a model-based approach and uses image-based cues, namely silhouettes and colour constancy. We show that, with the exception of a small set of parameter configurations, the cost function of our tracker has a well-behaved unique minimum. The convergence proof for the tracker relies on the convergence theory in stochastic approximation. We demonstrate that our tracker meets the sufficient conditions for stochastic approximation to hold locally. Experimental results on synthetic images generated from real hand motions show the feasibility of this approach.