Research On Vision-based Dynamichand Gesture Recognition

Hand gesture recognition(HGR) has always been one of the most important research fields of human-computer interaction(HCI) for the next generation. With the development of smart device, nowadays, more and more enterprises have set out to make a great investment to commercialize this technology.According to the sensor used, HGR can be classified into visual-based and non-visual-based. And the visual-based HGR can be futher classified into static hand gesture recognitioin(SHGR) and dynamic hand gesture recogni tion(DHGR). DHGR acquires user’s dynamic hand gesture video by a camera. Then it recognizes the user’s dynamic hand gesture order by a certain recognition algorithm. Finally,it gives a feedback according to the recognised order.The DHGR system described in this paper has been devided into four parts: skin segmentation, hand gesture modeling, dynamic hand gesture tracking and dynamic hand gesture classfication. Regarding the skin segmentation part, the paper has built a brightness-based muti-gaussian skin model for five color features. And it has come up with a skin-data collection method, by which we can collect skin samples for the skin modeling. Regarding the hand gesture modeling part, the paper has described the transition from hand gesture’s pixel expression to hand gesture’s geometric expression. This transition was effective and can be used to differentiate hand-skin data with non-hand-skin data(such as face-skin data). In addition, the geometric parameters has been taken as a feature vector for a particular hand gesture. Regarding the dynamic hand gesture tracking part, the paper has improved the Camshift algorithm and dynamic hand gesture has been tracked more effective. And regarding the dynamic hand gesture classification part, the paper has bound DTW with KNN. DTW has been used to calculate the similarity between dynamic hand gestures. And KNN has been used to choose the real order from various candidates by voting. In the end, we have achieves 81.31% accuracy of DHGR system and 0.235 s time consumption in single frame caculation.