Hand gesture recognition using RGB-D cues

In this paper, we propose a hand gesture recognition method in the clutter background by fusing the RGB-D cues. Since the hand localization is the key issue, we propose a coarse-to-fine procedure to detect hand accurately, which combines the statistic skin model using color information with depth prior knowledge. By detecting the skin candidate regions on the color image with Gaussian Mixture Model (GMM) skin model, hand region is obtained by compounding the depth information with the assumption that hands are at the closest position to the camera in all skin regions. Then, a new descriptor based on saliency point is used to represent the binary hand properly. A new hand model containing the wrist is proposed and the gesture recognition based on special points is applied. The experiment results demonstrate that our method performs better than NMI and moment based methods with a 96.2% recognition rate.