基于PCA-HOG与LBP特征融合的静态手势识别方法研究

发布时间：2018-11-27 12:26

【摘要】：随着计算机技术不断的发展,出现了越来越多的人机交互方式。由于手势的直观性、自然性等特点,所以手势识别也成为了一种重要的人机交互方式(HCI)。但手势自身具有的多样性,以及在时空上的差异性等特点,使手势识别成为一个极富挑战性的多学科交叉的研究课题。如何能够快速而准确的识别出手势所表达的意义,成为了人们研究的重点。本文以兼顾实时性和提高手势的识别率为研究目标,设计实现一个基于计算机视觉的静态手势识别系统,完成对预定义的6种静态手势的识别。论文首先讨论了几种常见的图像预处理方法,以用来去除图像的噪声和增强图像的质量,并分别对梯度直方图(HOG特征)和支持向量机(SVM)进行了相关的介绍。由于手势的多样性以及图像背景的复杂性,本文选择单一特征最为强大的HOG特征。与其他特征相比,HOG特征对于手势图像的光线变化和小幅度旋转方面有较强的鲁棒性。将HOG特征与SVM结合起来,作为手势的识别算法。实验结果表明,HOG结合SVM的方法对手势识别有较好的分类效果。在对手势图像分类训练时,常用的HOG特征维数较高,包含大量的冗余信息,使得特征的提取算法较为复杂。为了克服这一不足,提出一种改进算法,引入了主成分析法(PCA)对HOG特征进行降维处理,形成PCA-HOG特征,并与LBP特征相融合形成新的PCA-HOG+LBP融合手势特征。该融合特征既有手势边缘梯度信息,又有纹理特征信息,能有效弥补单一HOG特征的不足,提高手势在遮挡情况下的识别率。最后用Jochen Triesch手势库中的手势图像对本文的识别算法进行验证。结果表明,基于PCA-HOG+LBP特征的识别算法在提高手势识别率的同时也能更好的保证实时性。最后,基于Microsoft Visual Studio 2010和Open CV环境搭建了手势识别的原型系统,设计并实现了一个小型手势识别系统。论述了该系统流程,关键模块的实现代码等内容,通过摄像头采集手势图并自制手势库完成测试,实验结果表明证明改进后的算法在本系统是具有可行性的。
[Abstract]:With the development of computer technology, more and more man-machine interaction methods appear. Gesture recognition has become an important human-computer interaction method (HCI). Because of the intuitive and natural features of gestures. But the diversity of gesture itself and the difference in time and space make gesture recognition a challenging interdisciplinary research topic. How to recognize the meaning of gestures quickly and accurately has become the focus of research. In this paper, a static gesture recognition system based on computer vision is designed and implemented with the aim of considering real-time and improving the recognition rate of gesture. The recognition of six predefined static gestures is accomplished. In this paper, several common image preprocessing methods are discussed to remove image noise and enhance image quality. Gradient histogram (HOG) and support vector machine (SVM) are introduced respectively. Due to the diversity of gestures and the complexity of image background, the single feature is chosen as the most powerful HOG feature in this paper. Compared with other features, HOG features are robust to light variation and small rotation of gesture images. HOG features are combined with SVM as gesture recognition algorithms. The experimental results show that the method of HOG combined with SVM has better classification effect for gesture recognition. In the training of gesture image classification, the commonly used HOG feature dimension is high and contains a lot of redundant information, which makes the feature extraction algorithm more complex. In order to overcome this shortcoming, an improved algorithm is proposed. The principal component analysis method (PCA) is introduced to reduce the dimension of HOG features to form PCA-HOG features, and to merge with LBP features to form new PCA-HOG LBP fusion gesture features. The fusion feature has both gradient information of gesture edge and texture feature information, which can effectively compensate for the deficiency of single HOG feature and improve the recognition rate of gesture in occlusion. Finally, the recognition algorithm of this paper is verified by the gesture image in Jochen Triesch gesture database. The results show that the recognition algorithm based on PCA-HOG LBP features not only improves the recognition rate of gestures, but also ensures better real-time performance. Finally, a prototype system of gesture recognition is built based on Microsoft Visual Studio 2010 and Open CV, and a small gesture recognition system is designed and implemented. The flow of the system and the code of the key module are discussed. The gesture diagram is collected by the camera and the hand gesture database is made to complete the test. The experimental results show that the improved algorithm is feasible in this system.
【学位授予单位】：兰州理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】