Abstract. This article proposed a new method which introduces human body structure into gesture recognition besides the traditional vision information. By establishing the target's skeleton model, the system could rule out background interference quickly from the real-time frames, then locate the target's joints accurately, and track the gestures automatically. This method applies a common monocular webcam for image input instead of expensive depth cameras or complex binocular cameras. It can build a vivid human model and simulate the target's gestures well. After a modified KNN classification which has self-learning ability, the relationships between these joints are analyzed, and the corresponding postures and movements are identified. This method avoids global scanning in each frame and simplifies the classification, thus the computation decreases sharply. The results show that the method has a good tolerance of complex background, it can solve the target missing well.

No abstract