Enhanced hand tracking using the k-means embedded particle filter with mean-shift vector re-sampling

In recent years, particle filters have been applied with great success to 2D and 3D tracking problems. We presents the tracking of two hands based on a statistical model using only a skin colour feature with particle filtering for gesture recognition. The tracking scheme employs the reliability measurement derived from the particle distribution which is used to adaptively weight the skin-pixel colour classification. Our approach chooses shift-vectors to re-weight the particle sample to improve accuracy and reduce the number of samples. The k-means algorithm is used to discriminate the split and merge between left and right hands in case they are close together. Experiments with a set of videos including the movement of two hands in sample and cluttered backgrounds show that adaptive use of our scheme provides improvement compared to use with auxiliary particle filter of the number of samples and accuracy.