Study On Vision-based Dynamic Hand Gesture Recognition Technique Under Complex Background

People have been using hand gestures to communicate with each other for a very long time. Hand gesture is one of the most important way people use to carry on messages. With the development of the computer technology, the focus of HCI (Human Computer Interaction) is moving onto human. Vision-based hand gesture recognition promises a much friendlier way of HCI. It stands for the cutting edge and tendency of this field. However, hand gestures are rich in diversities, multi-meanings and space-time varieties. Human hands themselves are complex non-grid objects, so this field is extremely challenging and interdisciplinary. The technology is still in the laboratory.Vision-based dynamic hand gesture recognition is composed of three parts. They are hand gesture segmentation, hand gesture eigenvalue selection and eigenvalue recognition. This paper accomplishes the following works, based on a predefined processing framework.As to the hand gesture segmentation part, this paper proposes a refined procedure to segment the gesture image. A method that combines Kalman filtering to determine the hand region position with skin color detection in HSV color space is applied for segmentation. Besides, auto white balance based on priori skin color information and movement analysis based on R-channel image difference are employed to extract hand gestures from complex background. Compared with the method based on RGB color model, it is proved to get a better result.As to the hand gesture eigenvalue selection part, on analyzing a simple topology based descriptors and normalized Fourier descriptors, this paper proposed a new kind of rotation sensitive eigenvalue descriptors which is based on edge sampling statistics. The experiment proves that this kind of descriptors is effective and easy processing under the assumption that the rotation of hands may result in different meaning in dynamic hand gesture recognition.The eigenvalue recognition part is composed with track recognition and track recognition with finger information. This paper employs HMM (Hidden Markov Model) that has had great success in speech signal processing to recognize gesture track to recognize hand tracks, which is proved to be effective and satisfying. As to the track recognition with finger information, this paper proposed a combination of key frame recognition and track recognition, whose average recognition rate is 90.3%. This result is satisfying and has practical value.