Research On Vision-based Real-time Static Hand Gesture Recognition

Hand Gestures constitute a way of communication constitute a way of communication between human in everyday lives,which have the advantages of natural and simple. With the rapid development of computer science, gesture recognition is a new generation of human-computer interaction. In recent years, with the popularity of the camera, Vision-based gesture recognition technology is one of the focused research directions in the field of human-computer interaction.Vision-based static gesture recognition technology obtain the gesture images through the digital image acquisition system.Generally, the system complete gesture recognition through three steps:gesture segmentation, gesture feature extraction and gesture classification. The result of vision-based gesture recognition is under the influence of the light, differences of gestures, gesture rotation and other factors. The technology of gesture recognition based on directional gradient histogram feature(HOG) and support vector machine(SVM) classifier is practical through the research and comparison of various static gesture recognition technology. The method of HOG features extraction processes the image in the sub blocks, so that the relationship between the local pixel points can be describe d. And the influence of a small scale translation and small angle rotation and the uneven illumination of the image of can be suppressed in a certain extent. HOG features combined with the SVM method has been mature in the application of detection of objects, especially for the pedestrian detection. To use the technology for a variety of gesture classification, it need to split the full range of gestures and limit the rotation rate. Focusing on those problems, this paper do a corresponding research and improvement.Firstly, this paper studies and analyzes several skin color segmentation methods, and selects the brightness-based muti-gaussian skin model, which makes the segmentation algorithm suitable for complex background and illumination changes. Then, this paper designs a method to avoid the interference of the exposed arm to the gesture segmentation. The method to avoid the arm interference based on wrist positioning. The algorithm using fingertip detection based on curvature and distance transformation to determine the direction of gestures.Then it search the wrist from the direction of the fingers based on width. Using this method can find the wrist and remove the arm region quickly and accurately, meanwhile the problem of large scale of the gestures can be resolved. Finally, the HOG features are extracted from the segmented images. Then a classifier which combines SVM and Softmax is designed to reduce the classification time.In this paper, we have done some experiments to verify the availability of the results of the research. The recognition rate of the hand gestures which are free from the arm and other factors is 94.5%, the accuracy rate of the gesture recognition Under the complex background is 90.7%, and the time on the PC platform is 0.17 s for one image.