【作者】 张凯； 葛文兵； 汪国平； 董士海；

【Author】 Zhang Kai +, Ge Wenbing, Wang Guoping, Dong Shihai(HCI & Multimedia Lab, School of Electronics Engineering and Computer Science, Peking University, 100871, China)

【机构】 北京大学信息科学技术学院人机交互与多媒体实验室；

【摘要】 本文提出了一种基于立体视觉的自然手势识别方法。本文方法有着如下特点:i)无穿戴自然手势交互,在满足平面近似的条件下,可任意扩充手势语义;ii)模板库小,仅仅需要存储每种手势的正面图像,便于高效匹配;iii)提出了一种新的立体匹配算法,将手势的视差图简化为平面,用以快速获取手势和手臂的三维信息;iv)可自动调整相机位置,得到手势的正面图像,用于模板匹配。实验结果表明,本文算法对各种朝向的手势都有着良好的支持,识别率高,强壮性好。更多还原

【Abstract】 Gesture recognition is one of the most active research area in human-computerinterface, and stereo matching for extraction of three-dimensional scene structure has alsobeen an intense area of research for decades. In this paper we combine the techniques of twoaspects and propose a stereo vision based hand gesture recognition algorithm by templatematching. Our algorithm is mainly based on two assumptions. The first is gesture’s planarhypothesis by which we neglect the self-occlusion of hand gesture and deem the gesture’sdisparity map as a plane. The second is conplane hypothesis by which we presume hand andarm are in the same plane. If we have got the gesture plane and the arm direction, we canrotate and translate the camera to get the gesture’s frontal image, then gesture semantic can berecognised by comparing the frontal image with each template image. There are fourcontributions of our paper. First, we present a wearless gesture recognition method and naturalgestures are expandable. Next, we use a compact gesture template for high efficientrecognition, including only gestures’ frontal images. As a third contribution, we use aplane-based stereo matching algorithm to get the 3D geometry infomation of hand plane andarm direction. Finally, gesture’s frontal image is retrieved by automaticly rotating andtranslating camera’s position. Experiments demonstrate that high recognition rates areachieved in various types and poses of gestures, which shows our method is strong andpromising.更多还原