Research On Dynamic Hand Gesture Recognition Based On Machine Vision

Machine-vision-based dynamic hand gesture recognition, which is natural and intuitive, has become a kind of extremely friendly manner of HCI and show a promise market of application. However, the randomness of dynamic gesture and a great number of disciplines being involved makes the gesture difficult to be recognized exactly and efficiently. Moreover, the existing hand gesture recognition system is of low accuracy,low robustness and also inefficiency since it contains a serial of difficulties including the determination for the start or end point of a gesture trajectory, effective recognition for the similar gestures and the large variability of a gesture trajectory.Two important issues in a gesture recognition system are studied and a 3D trajectories based dynamic hand gesture recognition system is designed in the paper.Firstly, how to determine the start and end point of the gesture trajectory, that is effective trajectory determining. Other hand, how to improve the recognition rate of the similar gestures.The existing methods for effective trajectory determining make use of the moving speed and spatial location information and so on by detecting the moving information of the performing hand. The robustness and stability of these existing method are low, in order to solve which, a new method is proposed in the paper. The proposed method use the closed and open state of the performing fingers to realize effective trajectory determining. When the fingers are closed, the system is waiting for activating the gesture capture module. Once the open fingers are detected, our system begins to sample the moving hand and record the 3D coordinates information. The proposed method is good since the open or closed state is less dependent on human and our capture device here is sensitive to these two states. The experimental results show recognition with 94.2% for the proposed method, which shows efficiency, high robustness and stability of the proposed method.The existing methods for similar gesture recognition are time consuming and the accuracy is low, as for improving which, the non-similar and the similar gestures are separated to be classified. Only an HMM-based recognition system is used to recognizethe non-similar gestures. While for the similar gestures, four secondary features are extracted further for the input gesture and an enhanced recognition is implemented. In a word, vast non-similar gestures are recognized efficiently and exactly and the recognition rate for the similar gestures is improved significantly by the proposed enhanced recognition. The experimental results show that our proposed system gains compromise between recognition accuracy and speed and achieves high average accuracy with95.17% which is higher than the multi-feature-based and HMM-FNN-based methods.