基于Kinect的手势识别及其在场景驱动中的应用

发布时间：2019-02-16 08:47

【摘要】：在用户界面研究中,人机交互技术是当前发展最迅速的技术之一,研究人员予以特别重视。它是一门综合学科,与认知学、人机工程学、心理学等学科领域有着密切的联系。作为人机交互中重要的一部分,手势识别一直以来被众多研究者重视。特别是近几年,随着微软公司的Kinect的出现,符合人机交流习惯的手势识别交互技术的研究变得非常活跃。按照手势动作分类,手势识别研究包括两部分:静态手势识别及动态手势识别。本课题以微软公司提供的Kinect为手势动作的采集设备,对静态手势识别和动态手势识别的算法分别进行优化然后在虚拟场景中完成测试。首先,为了使手部区域分割更精确,提出一种新的手部区域分割算法。该算法通过计算躯干区域和手部区域的类间方差得到最佳分割阈值,从而提取到手部区域,再计算手部区域点密度最大的点得到掌心点,采用相应椭圆描述手掌区域的基础上结合相应坐标系将手部区域细分成手掌区域、指尖区域和手臂区域。其次,针对静态手势识别过程中利用单特征识别时准确率低的问题,提出一种基于多特征提取的手势识别算法。此算法首先提取指尖点到手掌中心点的距离、指尖点到手掌平面的距离和手掌区域三种不同的手势特征,然后应用一个多分类的支持向量机(SVM)分类器对静态手势进行分类,并在手势数据库中完成了算法验证。第三,针对动态手势识别过程中关节点获取不准确的问题,提出一种利用关节点可信度度量关节点有效性的算法。此算法通过计算关节点的行为可信度、运动学可信度和彩色图像可信度及其可信度的特征权重,可更准确获取动态手势的关节点,从而完成快速准确的动态手势识别。最后,在基于3ds Max和Unity 3d设计的三维虚拟场景中完成实时检测。结合静态手势和动态手势识别技术,设计包括开始、指向、转向、放缩、挥手及停止等手势动作,驱动虚拟场景完成相应功能的实时变化,验证了算法的有效性。
[Abstract]:In the research of user interface, human-computer interaction is one of the most rapidly developing technologies, and researchers pay special attention to it. It is a comprehensive subject and has close relation with cognitive science, ergonomics, psychology and so on. As an important part of human-computer interaction, gesture recognition has been paid attention to by many researchers. Especially in recent years, with the emergence of Microsoft Kinect, the research on gesture recognition and interaction technology, which accords with man-machine communication habits, has become very active. According to gesture classification, gesture recognition includes two parts: static gesture recognition and dynamic gesture recognition. In this paper, the Kinect provided by Microsoft is used as the acquisition device of gesture action. The algorithms of static gesture recognition and dynamic gesture recognition are optimized and tested in virtual scene. Firstly, in order to make hand region segmentation more accurate, a new hand region segmentation algorithm is proposed. The algorithm obtains the optimal segmentation threshold by calculating the variance between the torso region and the hand region, and then extracts the hand region, and then calculates the point with the highest density in the hand region to get the centerpoint. On the basis of describing the palm region with the corresponding ellipse, the hand region is subdivided into palm region, fingertip region and arm region in the corresponding coordinate system. Secondly, aiming at the problem of low accuracy when using single feature in static gesture recognition, a gesture recognition algorithm based on multi-feature extraction is proposed. The algorithm firstly extracts the distance from the fingertip to the center of the palm, the distance from the fingertip to the palm plane and three different gesture features in the palm area. Then, a multi-classification support vector machine (SVM) classifier is used to classify the static gestures. The algorithm is verified in the gesture database. Thirdly, aiming at the problem of inaccuracy of node acquisition in dynamic gesture recognition, an algorithm is proposed to measure the effectiveness of the node by using the reliability of the node. By calculating the behavioral credibility, kinematics credibility and the feature weights of the color image credibility, the algorithm can obtain the dynamic gesture nodes more accurately, so as to complete the fast and accurate dynamic gesture recognition. Finally, real-time detection is completed in a three-dimensional virtual scene based on 3ds Max and Unity 3D design. Combined with static gesture and dynamic gesture recognition technology, the design includes start, point, turn, drop, wave and stop gestures, drive the virtual scene to complete the corresponding real-time changes, and verify the effectiveness of the algorithm.
【学位授予单位】：中北大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】