改进Kinect手部行为识别系统的研究  

Improvement of hand behavior recognition on Kinect system

作　　者：曹国强 刘浩然 CAO Guoqiang;LIU Haoran(College of Mechanical and Electrical Engineering, Shenyang Aerospace University, Shenyang 110136, China)

机构地区：[1]沈阳航空航天大学机电工程学院,沈阳110136

出　　处：《黑龙江大学自然科学学报》2019年第1期109-114,共6页Journal of Natural Science of Heilongjiang University

基　　金：辽宁省自然科学基金资助项目(20170540711)

摘　　要：针对Kinect自带的动作模板库无法判断人体多种动作组合行为以及不同位置含义不同的问题,本文以人体手部行为为例,设立4种位置下4种基础手部动作,提出了在屏幕坐标系下构建运动特征模板,并提出建立一种DTW-NBC模型,将动态时间规整(Dynamic time warping, DTW)与朴素贝叶斯分类(Naive bayes classification, NBC)相结合进行模板训练与匹配,并对上述任意两种基础动作组合产生的行为在不同位置的发生进行识别。实验证明,该方法有效地区分了在不同位置相同的手势行为,改善了Kinect对同时发生的复合动作进行识别精确度较低的缺点,识别率达到88.6%,具备良好的稳健性和有效性。Aiming at the problem of the combinations of human behaviors failed to be recognized by the built-in motion template library of Kinect’s and the problem of different meanings for different positions, the human hand’s behavior is taken as an example to set up four basic hand’s actions for four different locations. A motion feature template is established in the screen coordinate system, and a DTW-NBC model is put forward by the combination of Dynamic time warping(DTW) and Naive Bayes classification(NBC) to train and match the template, which will identify the combinations of any two basic actions at different locations. Experiments show that the method can distinguish the same behavior at different locations effectively and improve the accuracy of recognizing combined motions on Kinect with the recognition rate reaching 88.6%, so as to achieve good stability and effectiveness at the same time.

关 键 词：KINECT 行为识别 动态时间规整 朴素贝叶斯分类 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]