基于动态手势控制的交互式体三维显示  

Interactive Volumetric Three-dimensional Display Based on Dynamic Hand Gesture Control

作　　者：潘文平[1] 沈春林[1] 张赵行[2] 邢建芳[1] 

机构地区：[1]南京航空航天大学自动化学院,南京210016 [2]中国科学院自动化研究所,北京100190

出　　处：《光电工程》2010年第12期88-96,共9页Opto-Electronic Engineering

基　　金：国家863资助项目(2007AA01Z338)

摘　　要：研究了动态手势的跟踪和识别算法,实现了基于动态手势控制的交互式体三维显示。体三维显示可在真实的三维空间中呈现具有物理深度的三维图像,并且所显示图像可从360°范围内任意角度裸眼观看。采用基于视觉的动态手势控制,可从围绕显示系统的任意视角实现交互式体三维显示。首先从相邻帧帧差图像中检测特定的静态手势获得初始跟踪区域,并针对此区域建立手部肤色的分布模型。然后通过光流法更新跟踪区域内KLT特征点的位置,并按特征点群算法结合手部肤色的分布模型对特征点的位置进行修正。对于跟踪获得的动态手势轨迹,经过坐标变换并按八方向Freeman链码进行量化编码,获得手势观测序列。最后利用经过训练建立的手势隐马尔科夫模型库,实现动态手势的识别。将动态手势控制应用到交互式体三维显示并进行了实验。实验结果表明,手势跟踪和识别算法准确而快速,对不同手势的平均识别正确率超过93%,平均响应时间小于40ms,能从任意视角较好地实现交互式体三维显示。Tracking and recognition algorithms for dynamic hand gestures have been studied and interactive volumetric three-dimensional （3D） display based on dynamic hand gesture control has been implemented. Volumetric 3D display can present images with physical depth cues in the true 3D space, where images can be viewed from almost any angle without any special eyewear. With vision-based hand gesture control, viewers can interact with volumetric 3D display from any viewpoint. Firstly, tracking region containing a special static hand posture is initialized from frame-differencing images. Meanwhile, a skin color model is built based on the region. Then positions of KLT features in the initial region are updated using optical flow tracking method and adjusted with the ＂flock of features＂ algorithm and the learned skin color model. Thereafter, through coordinate transformation and encoding according to eight-directional Freeman chain code, gesture observation sequence is generated for recognition. Finally, a pre-specified gesture symbol is decided by Hidden Markov Model （HMM） evaluation using a trained gesture HMM set. Experiments are conducted for the interactive volumetric display based on dynamic hand gesture control. Experimental results show that with these gesture tracking and recognition algorithms, the average recognition rate is high up to 93%, the average response time is less than 40 ms, and interactions can be well performed from arbitrary viewpoint.

关 键 词：隐马尔科夫模型 手势识别 体三维显示 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] TN141[自动化与计算机技术—计算机科学与技术]