A robust, self-initialising, real-time hand tracker

In this paper, we present a robust system for visual tracking of hands as a user interface alternative. Our markerless tracker is meant as a replacement for expensive VR tracking equipment using cheap off-the-shelf components. We perform a Bayesian simulation with arbitrary degree-of-freedom models of the human hand using a modified particle filter. To counter the jitter from the stochastic simulation we run a linear predictive filter on the estimates. Basic gesture recognition is possible through the simultaneous use of multiple model configurations. Test results show that our system is robust to missing data and that it can initialise and reinitialise automatically. Real-time operation is possible without significant optimisation efforts.