基于视觉的手势识别及其应用研究

发布时间：2018-02-24 13:48

本文关键词： 手势识别手势分割 Hu矩人工神经网络手势跟踪　出处：《南京理工大学》2017年硕士论文　论文类型：学位论文

【摘要】：随着计算机技术的发展,人与机器交互的重要性逐渐凸显。最初的人机交互通过键盘输入命令行进行,之后出现鼠标、触摸屏等方式。而以手势作为人机交互的输入,显得更加自然与便捷,更能体现出以人为中心的思想。基于视觉的手势识别有广泛的应用前景,可以用于安全驾驶、商务办公、手语识别、娱乐设施等,为人们的日常生活带来诸多方便,这些应用前景也使得许多国内外学者致力于这个领域的研究。本文针对基于视觉的手势识别,做了以下工作:首先,在手势分割方面,对摄像头采集的视频帧,采用混合滤波和直方图均衡化,提高了图像的对比度;采用基于肤色方法,使用YCrCb色彩空间对静态手势进行了分割;针对露出胳膊的情况,使用几何方法,去除了将手臂部分,最后得到仅包含手掌部分图像。其次,在静态手势识别方面,本文采用了手势图像的Hu矩特征作为手势特征,使用BP神经网络对其训练,得到静态手势分类器。对自定义的9种手势,得到95%以上的识别率,在行业内处于较高水平。然后,在动态手势识别方面,本文结合前文静态手势识别结果,对现有的TLD跟踪算法进行了改进,实现了对手势有效、实时的跟踪;采用隐马尔科夫模型,对6种常见的手势轨迹进行识别并取得较好的效果。最后,在上文手势识别研究的基础上,本文结合两个具体应用背景,开发了两个应用程序,对手势识别的应用进行了研究。第一个应用为"猜拳游戏",通过双摄像头分别识别两用户的静态手势,将其进行对比并输出对比结果;第二个应用为手势控制的音乐播放器,通过手势控制播放器的开启、音乐的音量和音乐切换等功能。
[Abstract]:With the development of computer technology, the importance of human-machine interaction becomes more and more important. At first, human-computer interaction is carried out by keyboard input command line, then by mouse, touch screen and so on. It is more natural and convenient, and can reflect the human-centered idea. Visual gesture recognition has a wide application prospect, can be used in safe driving, business office, sign language recognition, entertainment facilities, etc. It brings many convenience for people's daily life, and these application prospects also make many scholars at home and abroad devote themselves to the research in this field. In this paper, the following work is done for visual gesture recognition: first, in the aspect of gesture segmentation, For the video frame captured by the camera, mixed filtering and histogram equalization are used to improve the contrast of the image. The color space of YCrCb is used to segment the static gesture based on the skin color method. The geometric method is used to remove the arm, and finally only the palm image is obtained. Secondly, in static gesture recognition, the Hu moment feature of the gesture image is used as the gesture feature, and the BP neural network is used to train it. A static gesture classifier is obtained. The recognition rate of more than 95% is obtained for 9 kinds of gestures defined by ourselves, which is at a high level in the industry. Then, in the aspect of dynamic gesture recognition, this paper combines the results of static gesture recognition with the previous results. The existing TLD tracking algorithm is improved to realize the effective and real-time tracking of gestures, and six common gesture tracks are identified with hidden Markov model. Finally, a better result is obtained. On the basis of the research of gesture recognition above, two application programs are developed in this paper in combination with two specific application backgrounds. The application of gesture recognition is studied in this paper. The first application is "guessing boxing game", which recognizes the static gestures of two users through two cameras, compares them and outputs the contrast results. The second application is a music player controlled by gesture. Control the player by hand gesture, music volume and music switching functions.
【学位授予单位】：南京理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】