Surface electromyography (sEMG) is derived from human skeletal muscle and can be employed as an accurate signal to reflect the body's muscle state. Because of its non-invasiveness and convenience, it is commonly used in many fields such as pattern recognition-based medical rehabilitation and intelligent humanoid robots. In view of the fact that sEMG signals are weak and easy to be interference by various noise, this paper proposes a method of sEMG preprocessing based on wavelet threshold method to eliminate the noise. At the same time, the RMS value, the waveform length feature, and the nonlinear feature sample entropy are extracted to form the feature vector for subsequent pattern recognition. Finally, sEMG signals of 9 common hand movements were used to establish the gesture recognition model based on SVM. The recognition rate reached over 99%, which verified the effectiveness of the proposed method.