Visual Tracking of Highly Articulated Objects Using Massively Parallel Processors by Dennis

Hand gesture recognition has the potential of simplifying human computer interactions. However, the human hand is a highly articulated object, capable of taking on many different appearances. In this work, we consider an analysis by synthesis approach to this difficult tracking problem. We attempt to overcome the vast amount of computation required by implementing the algorithm on commodity graphics processing units (GPUs). We also collect a lengthy sequence of hand motions from five cameras in order to train and test our algorithm. We show that to achieve good tracking performance, it is important to understand the way that the hand moves. It is of secondary importance to have a good estimate of the hand shape and to be able to process the frames as quickly as possible. Under heavily controlled circumstances, we are able to achieve full tracking accuracy.