Monocular Vision-based Real-time Hand Gesture Recognition System

With the development of the computer techniques, gesture recognition is becoming one of the key techniques of human-computer interaction technology. It is a hot research topic in the fields of image processing, pattern recognition, computer vision etc. However, hand gestures recognition is an extremely challenging inter-disciplinary project, due to two reasons: firstly, hand gestures are rich in diversities, multi-meanings, and space-time varieties; secondly,human hands are complex non-rigid objects. This paper presents a vision-based hand gestures recognition algorithm from points of pre-processing, feature extraction and recognition of hand gestures image based on the national 863 programs"anthropomorphic human-machine interaction system based on gesture".This paper studies the gesture recognition related fields, designs and implements a static monocular vision-based gesture recognition system. This system can capture and recognize 14 common static gesture in real-time and can control input method edit. This system has high recognition accuracy and real-time characteristics. The system is mainly divided into three parts. First, preprocession of the original hand gesture image: the experimental results show that the value of human skin color varies in a narrow range and it has obvious property of skin color clustering. Accordingly, this paper uses hsv color space for gesture region segmentation. After segmenting gesture region, the system gets the edge through noise smoothing and Laplacian edge extraction; Second, extraxction of hand gesture feature, in this part the system extracts Hu moment invariant, hand gesture area feature and Fourier descriptor; Third, the real-time procession to the video data stream, in this part the system compares the effect of Bayes, MLP machine learning method, and at last the system selects MLP as classifier.The experimental results show that the gesture recognition method based on gesture area feature, Hu moment invariant feature, Fourier descriptor feature and MLP classifier in this paper has a higher recognition accuracy(97.4%) which is consistent with the design criteria of high recognition rate and real-time processing.