基于视觉的手势识别方法研究

发布时间：2018-05-25 16:54

本文选题：手势识别 + Kinect　；参考：《兰州交通大学》2017年硕士论文

【摘要】：计算机科技进步的成果已经深入到人们生活的方方面面,而联系起人与机器之间沟通桥梁的正是人机交互技术(Human Computer Interaction,HCI)。在以用户体验为主导的今天,基于视觉的手势交互以其便捷、自然和友好等优势,逐渐成为人机交互技术研究的重要分支。随着Kinect等具有深度信息捕捉功能的传感器的出现,基于视觉的手势识别研究有了新的发展方向。由于手势形态在空间和时间上灵活可变,使得基于视觉的手势识别在实际应用中有着巨大的发展潜力。本文基于Kinect传感器就复杂背景下静态手型和动态手指轨迹相结合的手势识别方法开展研究,主要在手势分割、指尖检测、手势特征提取、手势识别和人机交互手势设计方面进行了一系列的研究与实验。本文主要工作如下:(1)手势分割部分,针对背景中脸部和类肤色区域对手势分割效果影响较大的问题,首先利用YCbCr椭圆肤色模型分割出肤色区域,并对得到图像进行腐蚀处理以去除噪声;其次利用Kinect传感器所获的深度信息对肤色区域进行投影,采用自适应深度阈值方法分割出手势。(2)指尖定位部分,针对现有基于轮廓曲率定位指尖存在的问题,本文引入基于凸包的指尖定位方法,首先获得手势轮廓并计算其近似多边形;其次计算轮廓凸包顶点从而获得指尖候选点;最后通过相邻凸缺陷最深点与凸顶点夹角和凸缺陷深度筛选出指尖,并获取各个指尖位置信息从而得到指尖的运动特征。(3)手势特征提取部分,为了使手势交互更为自然并且保证识别准确率,在提取静态手势特征方面,本文选择了手势的结构特征(指尖个数,手势轮廓周长面积比)和统计特征(Hu距的前四阶距)组成特征向量。(4)手势识别部分,本文构建了支持向量机的多值分类器,对本文所设计的交互手势的静态手势部分进行识别,结合手指运动特征最终识别出交互手势并触发操作。经实验分析,本文设计的手势交互方式自然灵活且识别率高。
[Abstract]:The achievements of computer science and technology progress have penetrated into every aspect of people's life, and it is the human-computer interaction technology, Human Computer interaction, that links the bridge between human and machine. Nowadays, with the user experience as the dominant factor, visual gesture interaction has become an important branch of human-computer interaction technology for its advantages of convenience, nature and friendliness. With the appearance of sensors with depth information capture function such as Kinect, the research of hand gesture recognition based on vision has a new development direction. Because gesture forms are flexible in space and time, visual gesture recognition has great potential in practical applications. Based on the Kinect sensor, this paper studies the hand gesture recognition method based on the combination of static hand type and dynamic finger trajectory in complex background, mainly in gesture segmentation, fingertip detection, gesture feature extraction, etc. A series of researches and experiments have been carried out on gesture recognition and human-computer interaction gesture design. The main work of this paper is as follows: (1) gesture segmentation. Aiming at the problem that the facial and skin-like regions in the background have a great influence on the gesture segmentation effect, we first use the YCbCr elliptical skin color model to segment the skin color region. The image is corrupted to remove noise. Secondly, the depth information obtained by Kinect sensor is used to project the skin color area, and the finger tip location part is segmented by adaptive depth threshold method. Aiming at the problems existing in the existing fingertips location based on contour curvature, this paper introduces a fingertip localization method based on convex hull. Firstly, the gesture contour is obtained and its approximate polygon is calculated; secondly, the contour convex hull vertex is calculated to obtain the finger tip candidate points. Finally, the finger tip is screened by the angle between the deepest point of the adjacent convex defect and the convex vertex and the depth of the convex defect, and the position information of each finger tip is obtained so as to obtain the motion feature of the fingertip. In order to make gesture interaction more natural and ensure recognition accuracy, this paper chooses the structural feature of gesture (number of fingertips) to extract static gesture features. In this paper, a multi-valued classifier of support vector machine is constructed, and the static gesture part of the interactive gesture designed in this paper is recognized. Combined with finger motion features, the interactive gesture is finally recognized and the operation is triggered. Experimental analysis shows that the gesture interaction method designed in this paper is naturally flexible and has a high recognition rate.
【学位授予单位】：兰州交通大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】