基于几何特征的手势识别算法研究

【作者】 何阳清；

【导师】 葛元；

【作者基本信息】 上海海事大学 ， 计算机软件与理论， 2004， 硕士

【摘要】 手势是人们生活当中一种自然而直观的人际交流模式，随着计算机技术的发展和人机交互逐渐向以人为中心转移，对手势识别的研究也逐渐成为人们研究的热点。然而，由于手势本身具有的多样性、多义性、以及时间和空间上的差异性等特点，加之人手是复杂变形体及视觉本身的不适定性，因此基于视觉的手势识别是一个极富挑战性的多学科交叉研究课题。手势分为动态手势和静态手势，动态手势定义为手运动的轨迹，而静态手势强调通过手型传递一定的意义。本文结合上海市自然科学基金资助课题“手势识别和合成算法”，对静态的手势识别算法进行研究。 手势识别的过程大致分为三个部分，手势图像预处理、手势图像特征提取和识别。在手势图像预处理部分，对已经被标准化的手势图像(大小为128*128像素的bmp格式的灰度图)，根据需要采用局部平均法对图像进行平滑，然后对图像采用拉普拉斯算子进行锐化，再对图像采用最大方差法进行二值化，最后用八方向邻域搜索法对二值化图像做轮廓提取。 在手势特征提取和识别部分，本文提出了两种基于手势图像几何特征的方法：HDC提取关键点的识别算法及应用几何矩和canny边缘检测结合的识别算法。在HDC提取关键点的识别算法中，提出一种提取手势轮廓曲线关键点对手势进行识别的算法。手势图像经过二值化后，提取其轮廓。将图像的轮廓看成一条曲线，应用层次离散相关原理，以一个内核对曲线进行多次平滑，得到曲线的尺度空间，再通过跟踪曲线在尺度空间中的运动找出手势轮廓的关键点。最后通过最小距离法进行识别。在应用几何矩和canny边缘检测结合的识别算法中，提出一种结合几何矩和边缘检测的手势识别算法。手势图像经过二值化处理后，提取手势图像的几何矩特征，取出几何矩特征七个特征分量中的四个分量，形成手势的几何矩特征向量。在灰度图基础上直接检测图像的边缘，利用直方图表示图像的边界方向特征。最后，通过设定两个特征的权重来计算图像间的距离，对30个字母手势进行识别。 最后，HDC提取关键点的识别算法在实验中对30个手势进行识别，识别率为83.3％。应用几何矩和canny边缘检测结合的识别算法结合了两种图像特征的优点，在实验中识别率为91.3％。更多还原

【Abstract】 Hand gestures play a natural and intuitive communication mode for all human dialogs. With the development of computer technology, HCI(Human Computer Interaction) is advancing and human is becoming the center in HCI. So a growing number of researcher are concerning the study on hand gesture recognition. However, vision-based recognition of hand gestures is an extremely challenging interdisciplinary project due to following three reasons: (1) hand gestures are rich in diversities, multi-meanings, and space-time varieties; (2) human hands are complex non-rigid objects; (3) computer vision itself is an ill-posed problem. Hand gestures include dynamic hand gestures, whose meanings are based on the track of the motion of hands, and static hand gestures in which the shape of hand gesture is used to express the meaning. This paper, as a part of the researching subject, the algorithm of hand gesture recognition and synthesizing, is supported by shanghai nature and science fund, tries to perform study on static hand gesture recognition.Hand gesture recognition is composed of three parts, preprocessing hand gesture images, extracting image features and recognition. During the preprocessing, image smoothing, then image sharpening are performed on standard hand gesture images (128*128 pixels gray bmp image). Finally, binary image is extracted and contour is detected by means of 8-connected boundary tracking when necessary.In the part of feature extraction and recognition, this paper presents two methods based on geometric features: Algorithm by Using HDC for Feature Pixels and Algorithm Based on Invariant Moment and Edge Detection. In the Algorithm by Using HDC for Feature Pixels, the contour of hand gesture, which will be regarded as a curve, is extracted after preprocessing. Then a scale space of the curve is created by the application of the hierarchical discrete correlation. Anew method which is based on the motion of the curve through scale space is proposed for feature detection. Finally, gesture patterns are recognized by means of minimal distance of feature pixels. In Algorithm Based on Invariant Moment and Edge Detection, an algorithm based on two featuresof invariant moment and edge detection is presented. After preprocessing, binary image is obtained and then 4 from 7 invariant moments are extracted. By edge detection, histogram is formed to describe the edge information. Finally, the recognition is performed on 30 letter gestures by computing distance, in which different coefficients are set to these two features.The recognition rate of is 83.3% in Algorithm by Using HDC for Feature Pixels by performing recognition on 30 hand gestures. In Algorithm Based on Invariant Moment and Edge Detection, the recognition rate of 91.3% is achieved.Yangqing He(Computer Software and Theory)Directed by-. Yuan Ge更多还原