A static hand gesture recognition system for human-robot interaction is proposed and realized. Hand area is segmented based on skin color model, then the hand contour is described by use of Fourier descriptor. LS-SVM is used as classifier, which is effective to small sample and has limited generalized error. A new incremental training algorithm for LS-SVM is proposed, which avoids the time-cost computation for matrix inverse. DAGSVM algorithm is used to combine two-class LS-SVM to realize multi-class gesture recognition. The recognition rate of 26 alphabetical gestures is the highest (93.62%) comparing with multi-layer perception, radial basis function network etc.