基于视觉的静态手势识别中关键技术的研究

发布时间：2018-10-11 18:28

【摘要】：在信息技术高速发展的今天,计算机作为一项伟大的发明,正深刻地影响着人们生活的方方面面。以生物识别技术为基础的自然人机交互技术作为计算机技术的一项重要应用,与人们的日常生活息息相关。基于计算机视觉的生物识别技术是指利用计算机技术对图像或视频进行处理,通过提取人体特有的生物特征,实现对生物体的识别,该技术正在成为人工智能领域的一大研究热点。利用生物识别技术进行人机交互相比传统的技术具有便利性、唯一性等优点。常用的生物特征包括人脸、指纹、虹膜和手势等,其中手势特征相比其他生物特征具有生动、自然和信息量丰富的特点。但是,由于人手具有不确定性和多重性,手势识别技术还有许多问题有待解决,因此手势识别正在成为人机交互领域研究的热点及难点。手势识别系统主要包括三个部分:图像预处理、特征提取和分类识别。本文主要研究了基于视觉的静态手势识别的相关算法,重点研究了特征提取算法与分类识别算法。针对这两部分内容,本论文主要做了以下工作:第一,详细研究了经典的特征提取算法和分类识别算法,并对它们的算法原理、算法步骤以及优缺点进行了细致的归纳总结。第二,针对基本的局部二值模式(Local Binary Patterns,LBP)算法识别率偏低,提取的特征维数偏大的缺点,本文提出了一种基于多邻域加权融合的局部二值模式算法,该算法是对基本LBP算法的一种改进。分别利用不同的处理策略由每一个中心像素点外两层的邻域点计算得到两幅LBP编码图像,并对它们进行统计得到两幅256维的直方图,然后将这两幅256维的直方图均匀量化为32维,最后将这两个32维的直方图进行加权融合得到一个32维的直方图作为最终的特征向量。通过在手势数据库上的实验结果证明,我们所提出的改进算法能够在提高手势识别率的同时大幅度地降低特征维数,从而提高运算速度。第三,深入研究了非负矩阵分解(Non-Negative Matrix Factorization,NMF)算法与压缩感知(Compressive Sensing,CS)算法,并利用这两种算法设计了一个手势识别系统。首先利用NMF算法将原始高维的图像向量投影到低维的子空间中得到低维的特征向量,然后利用CS算法所设计的分类器对得到的低维特征向量进行分类,得到手势识别的结果。通过一系列的实验证明,利用CS算法设计的分类器相比其他几种常用的分类器能够获得更高的手势识别率,同时具有更好的抗遮挡能力。另一方面,相较于主成分分析法(Principal Components Analysis,PCA),NMF算法对于遮挡具有更好的鲁棒性。
[Abstract]:With the rapid development of information technology, computer, as a great invention, is deeply affecting every aspect of people's life. As an important application of computer technology, natural human-computer interaction technology based on biometrics is closely related to people's daily life. Biometric recognition technology based on computer vision refers to the use of computer technology to process images or video, through the extraction of the unique biological characteristics of the human body, the realization of biological recognition. This technology is becoming a research hotspot in the field of artificial intelligence. Compared with the traditional technology, biometrics has the advantages of convenience and uniqueness. The commonly used biometric features include face fingerprint iris and gesture. Gesture features are more vivid natural and informative than other biometric features. However, due to the uncertainty and multiplicity of human hand, there are still many problems to be solved in hand gesture recognition technology, so gesture recognition is becoming a hot and difficult point in the field of human-computer interaction. Gesture recognition system consists of three parts: image preprocessing, feature extraction and classification recognition. This paper mainly studies the algorithms of static gesture recognition based on vision, especially the feature extraction algorithm and classification recognition algorithm. For these two parts, this paper mainly does the following work: first, the classic feature extraction algorithm and classification recognition algorithm are studied in detail, and their algorithm principle, algorithm steps, advantages and disadvantages are summarized in detail. Secondly, in view of the low recognition rate and large feature dimension of the basic local binary pattern (Local Binary Patterns,LBP) algorithm, a local binary pattern algorithm based on multi-neighborhood weighted fusion is proposed in this paper. This algorithm is an improvement on the basic LBP algorithm. Using different processing strategies, two LBP coded images are calculated from two adjacent points outside each central pixel, and two 256-dimensional histograms are obtained by statistical analysis. Then the two 256-dimensional histograms are uniformly quantized to 32-dimensional. Finally, the two 32-dimensional histograms are weighted and fused to obtain a 32-dimensional histogram as the final feature vector. The experimental results on the gesture database show that the improved algorithm can greatly reduce the feature dimension while increasing the recognition rate of the gesture, thus increasing the operation speed. Thirdly, the non-negative matrix decomposition (Non-Negative Matrix Factorization,NMF) algorithm and the compression sensing (Compressive Sensing,CS) algorithm are studied, and a gesture recognition system is designed using these two algorithms. First, the original high-dimensional image vector is projected into the low-dimensional subspace by NMF algorithm, and then the low-dimensional feature vector is classified by the classifier designed by the CS algorithm, and the result of gesture recognition is obtained. Through a series of experiments, it is proved that the classifier designed by CS algorithm can obtain higher gesture recognition rate and better ability to resist occlusion than other classifiers. On the other hand, (Principal Components Analysis,PCA), NMF algorithm is more robust to occlusion than principal component analysis (PCA).
【学位授予单位】：山东大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】