基于RGB-D信息的动态手势识别方法

Dynamic gesture recognition approach based on RGB-D information

针对现有的动态手势识别方法在复杂环境下,易受无关肤色、光照变化等因素的影响,识别率低、实时性差等问题进行了研究,提出一种动态手势识别方法。该方法首先利用K-均值聚类算法和YCr'Cb'(由YCr Cb变换得到)椭圆肤色模型对RGB-D图像完成手势分割;然后将深度信息引入到传统卡尔曼滤波算法中,作为其跟踪参数之一,并在跟踪过程中对检测范围进行加窗处理;最后结合快速动态时间规整算法和突出关键特征点的思想,改进传统动态时间规整算法,并利用改进后的动态时间规整算法完成手势识别。实验表明:提出的手势识别方法在复杂背景下的识别率较高(96.8±1.5%),实时性较好(识别时间1.86±0.02 ms)。

In order to solve the problems that the existing dynamic hand gesture recognition method in a complex environment has a low recognition rate and poor real-time performance,and would be affected easily by the factors such as irrelevant skin color,illumination change,this paper proposed a dynamic gesture recognition algorithm. Firstly,this paper introduced the Kmeans clustering algorithm and the YCr＇Cb＇ ellipse skin color model to segment the RGB-D image. Then,it introduced the depth information into the traditional Kalman filtering algorithm,which was regarded as one of the tracking parameters to process the detection range for adding windows in the tracking process. At last,it improved the traditional dynamic time warping algorithm by combining the fast dynamic time warping algorithm and the key feature points,and recognized the gesture with the improved dynamic time warping algorithm. The experimental results show that the proposed method has a relatively high recognition rate（96. 8 ＋ 1. 5%） in complex background and good real-time performance（ the recognition time is 1. 86 ＋0. 02 ms）.