Visual hand gesture recognition with convolution neural network

Hand gestures are a type of communication that is multifaceted in a number of ways and they provide an attractive alternative to the cumbersome interface devices used for human-computer interaction (HCI). However, there are still limitations regarding its usage in unfavorable live situations where hand gestures variation, illumination change or background complexity are an issue. Therefore, this paper propose a convolution neural network (CNN) method to reduce the difficulty of gestures recognition from a camera image. To achieve the robustness performance, the skin model and background subtraction are applied to obtain the training and testing data for the CNN. Since the light condition seriously affects the skin color, we adopt a simple Gaussian skin color model to robustly filter out non-skin colors of an image. In addition, it also employs elastic distortions to obtain lager database for more effective training and reduce potential overfitting. Experimental evaluation achieves an average correct classification rate of 93.8%, which shows the feasibility and reliability of the method.