Research On Vision Based Hand Gesture Recognition Technology

Human-Computer Interaction (HCI) has become an important part of our daily life as a result of the popularization of computers and the Internet. As computers are becoming more and more powerful, people begin to explore natural HCI methods. Vision based hand gesture recognition, which conforms to human communication habits, can provide long-distance and contactless interaction with computers. Therefore, it has become the focus of research in the field of natural HCI.This paper firstly introduces the background of hand gesture recognition and reviews the research of vision based hand gesture recognition in recent years. Secondly, a multi-cue based hand localization method is presented, which makes a combination of skin color detection and motion detection so as to improve localization results. Skin color pixels are detected by a proposed adaptive skin color model, which is updated using the previous results of skin detection and motion detection at each frame. Motion is detected by background subtraction while the background is updated thorough the running average method. The results of skin color detection and motion detection are fused to generate hand gesture candidates. Thirdly, a static hand gesture recognition method is proposed based on histograms of orientated gradients (HOG) and PCA-LDA. HOG features are extracted from normalized hand gesture images and then projected to low-dimensional subspace by PCA-LDA.Classification or recognition results are given in the subspace by the nearest neighbor method.The method achieves a recognition rate of more than 95% on the testing set and outperforms related static hand gesture recognition methods.Finally,the hand lacalization method and hand gesture recognition method are combined to make a pretotype system to recognize hand gestures in video. The performance of the pretotype system is evaluated by its ROC curve.Considering the possible orientation of the hand in natural scene,oreented training samples are added to train the gesture classifier,which brings obvious improvement in the classifier performance. A continuous hand gesture recognition method is proposed, which take the advantage of temporal correclation to segment and recognize continuous hand gestures. Using this method,93% of the defined contious hand gesture segments are classified correctly while 83% of the undefined segments are correctly detected.