基于Kinect的复杂手势识别技术研究

发布时间：2018-04-20 00:25

本文选题：Kinect + 手型分割　；参考：《南京理工大学》2017年硕士论文

【摘要】：随着人机交互技术的不断发展,各种新奇的人机交互方式层出不穷,手势识别技术以其学习成本低、灵活性好、实用性强等特点,近年来成为研究的热点。基于视觉技术的手势识别受光照、噪声等因素的影响较大,限制了对手势识别技术的应用。Kinect传感器能够在获取二维图像的同时获得空间的三维深度信息,为手势识别的研究带来了新的方向,本文利用Kinect2.0深度传感器,对具有手型变化的动态复杂手势的进行识别。主要包括手型图像的分割、手型特征和手势运动特征的提取、手势的分类识别等步骤。首先对于手型图像的分割,采用了 Kinect骨骼跟踪技术和深度信息相结合的方法,有效的消除背景和光照对手型图像分割的影响。对获得的手型二值图像进行形态学的处理,并采用边缘跟踪算法来实现手型图像轮廓的提取。然后是手势特征的提取,具体包括静态手型的特征和运动轨迹特征。提出对静态手型轮廓提取Hu特征并利用K-means聚类算法进行特征编码的方法;对于运动轨迹提取方向角特征,并进行球面14方向的量化编码,得到方向角特征编码,手型特征编码和方向角特征编码分别组合起来得到手势的手型特征序列和轨迹方向角特征序列。最后对手势进行分类识别,将隐马尔科夫模型和朴素贝叶斯模型相结合,提出了一种HMM-NBC模型进行手势的训练与识别。对于自定义的10种动态手势,平均识别率达到了 88.4%。
[Abstract]:With the continuous development of human-computer interaction technology, a variety of novel human-computer interaction methods emerge in endlessly. Gesture recognition technology has become a hot research topic in recent years because of its low learning cost, good flexibility, strong practicability and so on. Gesture recognition based on visual technology is greatly affected by illumination, noise and other factors, which limits the application of gesture recognition technology. Kinect sensor can obtain three-dimensional depth information of space while obtaining two-dimensional images. It brings a new direction to the research of hand gesture recognition. In this paper, we use the Kinect2.0 depth sensor to recognize the dynamic and complex hand gesture with the change of hand shape. It mainly includes the segmentation of hand image, the extraction of hand shape feature and gesture motion feature, the recognition of hand gesture classification and so on. Firstly, Kinect bone tracking technique and depth information are used to effectively eliminate the influence of background and illumination on hand image segmentation. The obtained binary image is processed by morphology and edge tracking algorithm is used to extract the contour of the hand image. Then the gesture features are extracted, including static hand features and motion trajectory features. A method of extracting Hu features from static hand contours and using K-means clustering algorithm to encode features is proposed. Hand feature coding and directional angle feature coding are combined to obtain hand gesture feature sequence and trajectory direction angle characteristic sequence respectively. Finally, a HMM-NBC model is proposed to train and recognize gestures by combining hidden Markov model with naive Bayes model. For the 10 kinds of dynamic gestures, the average recognition rate is 88. 4%.
【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】