具有不同数目状态结点的HMMs在中国手语识别中的应用  

THE APPLICATION OF HMMs CONSISTING OF DIFFERENT NUMBER STATES IN CHINESE SIGN LANGUAGE RECOGNITION

作　　者：王春立[1] 高文[2] 

机构地区：[1]大连理工大学计算机科学与技术系,大连116023 [2]哈尔滨工业大学计算机科学与工程系,哈尔滨150001

出　　处：《计算机研究与发展》2001年第1期111-115,共5页Journal of Computer Research and Development

基　　金：国家自然科学基金!(6 978930 1);中国科学院百人工程计划基金资助

摘　　要：中国手语是中国聋人使用的语言 ,主要通过手势动作来表达一定的含义 .因而 ,手语识别问题是动态连续信号的识别问题 .目前大部分手语识别系统采用 HMMs(hidden Markov models)作为系统的识别技术 .由于各个词包含的基本手势数不同 ,若所有模型都由同样数目的状态结点构成会影响识别率 .而由人为每个词设置状态数又很难达到完全准确 ,所述系统使用一种基于动态规划的估计状态结点数的办法 ,并实现了基于具有不同状态数目的 HMM的训练及识别过程 ,实验结果表明 。Chinese sign language (CSL) is the language used by many deaf people in China. The signer expresses himself by gestures and hand motions. Therefore CSL recognition is a problem of dynamic continuous signal recognition. Most sign language recognition systems use HMMs(hidden Markov models) presently. But the numbers of basic gestures included in different words are variable, so the accuracy will be affected if all models consist of the same number of states. It is difficult to set the exact number of states manually. In this paper, an approach based on dynamic programming is proposed to estimate the number of the states, and a CSL training and recognition system are implemented based on HMMs consisting of different number of states. The result of the experiments shows that the speed and accuracy of CSL recognition are improved.

关 键 词：马尔可夫模型 动态规划 手语识别 语音识别 计算机 HMMs 

分 类 号：TN912.34[电子电信—通信与信息系统]