A Research On Hand Gesture Recognition Based On Skeletonized Method

As a new research direction of human-computer interaction, the vision-based hand-gesture recognition is of vital importance for its wide application. Aiming at solving several currently-existing problems in the field of the static hand-gesture recognition, this paper puts forward an effective skeletonized method to realize image recognition which solves the problems of stability and accuracy. The main research done in this paper is as follows:First and foremost, based on the skeletonized method of distance transformation a modified method is proposed to extract skeleton features. The first step of this method is fast and efficient image Euclidean distance calculation. Then, initial skeleton growth. Finally, based on the discrete skeleton curve evolution of pruning. The new method is greatly modified in three ways. One is that it reflects the precision of traditional distance transformation method. Secondly, skeleton growth algorithm based on traditional distance transformation realizes skeleton’s single pixel and connectivity. Thirdly, The skeleton pruning method based on discrete curve evolution optimizes the coarse skeleton by reducing boundary noise and keeping the main part of the vision.Second, this paper uses boundary Hu moments to describe skeleton features. It is in this way that the regional factors of the traditional regional Hu moments can be eliminated and the descriptor of boundary Hu moments with complete information of Hu moments is realized. This robust method can effectively describe the skeleton and guarantee the following categorization and recognition based on SVM.Experiments show that the skeleton extracted in this paper has high stability, strong controllability and good connectivity. Compared with other traditional skeletonized method, this one has a great improvement in the following aspects. As the good descriptor of skeleton features, skeleton description based on boundary Hu moments possesses excellent stability and high realization efficiency. The machine learning recognition method based on SVM produces a quite ideal result, and its effect is also much better than that of other classic machine learning classification algorithm.