基于惯性传感器的手势交互方法研究

发布时间：2018-05-26 01:09

本文选题：手势交互 + 样本聚类　；参考：《电子科技大学》2017年硕士论文

【摘要】：智能型的用户界面操控技术日益受到重视,手势交互的方式具有学习成本低、自然便捷和多样丰富的特点,可以为操作者提供更为直观、舒适的自然交互体验。传统的基于惯性传感的手势交互方法的研究焦点集中于如何使不依赖于个体的手势识别方法更具有个体鲁棒性,同时获得更快的动态响应。但并未深入考虑算法中样本集的正规性和有效性,在一定程度上影响算法的识别准确率。同时,当手势复杂且手势种类增多时,传统方法更容易受到手势信号中冗余信息及噪音信息的影响,造成手势类别的误判。针对传统方法的不足与劣势,为提高手势识别的准确率和降低运算复杂度,本文进行了算法改进。实验表明,本文方法的运算耗时较传统DTW算法减少25%至31%,整体平均识别准确率在96.7%至98.84%,明显优于其他传统算法。本文主要致力于以下三个方面的研究工作:1.针对传统方法中样本集构造问题,为改善样本选取的非正规性,本文提出一种基于CDTW算法的样本聚类训练方法。该方法不仅可以克服不同个体手势动作速度差异大的问题,还改善了样本字典的正规性和有效性。通过样本聚类获得的典型样本,在一定程度上压缩了样本集的大小,更重要的是包含了不同个体做各个不同手势时的典型特征,令该方法具有更强的个体适应能力。2.针对传统手势交互方法在识别过程中运算量大的问题,聚类样本的方法可以使这一问题得到一定改善,本文同时提出主轴分类思想,运算中测试手势序列只与主轴相同的模板进行匹配,能够有效减少在线模板匹配过程在整个手势交互系统中的时间复杂度,从而确定其所属类。3.针对传统方法更容易受到手势信号中冗余信息及噪音信息的影响从而造成手势类别的误判的问题,提出利用压缩传感方法作为手势交互中识别过程的手段。利用这种压缩降维的方法,既减少了运算量,又可以不失真地恢复手势信号,保留重要特征,进而提高算法识别率,得出识别结果。
[Abstract]:Intelligent user interface manipulation technology has been paid more and more attention. Gesture interaction has the characteristics of low learning cost, convenience and diversity, which can provide operators with a more intuitive and comfortable experience of natural interaction. Traditional gesture interaction methods based on inertial sensing focus on how to make individual independent gesture recognition methods more robust and obtain faster dynamic response. However, the normality and validity of the sample set in the algorithm are not considered deeply, and the recognition accuracy of the algorithm is affected to some extent. At the same time, when the gestures are complicated and the kinds of gestures increase, the traditional methods are more vulnerable to the redundant information and noise information in the gesture signals, resulting in the false judgment of gesture types. In order to improve the accuracy of gesture recognition and reduce the computational complexity, the algorithm is improved to overcome the shortcomings and disadvantages of traditional methods. Experimental results show that the computational time of this method is 25% to 31% less than that of the traditional DTW algorithm, and the overall average recognition accuracy is from 96.7% to 98.84%, which is obviously superior to other traditional algorithms. This paper is devoted to the following three aspects of research work: 1. In order to improve the informality of sample selection, a training method of sample clustering based on CDTW algorithm is proposed to solve the problem of sample set construction in traditional methods. This method can not only overcome the problem of different individuals' gesture speed, but also improve the regularity and effectiveness of sample dictionary. The typical samples obtained by sample clustering can compress the size of the sample set to a certain extent, and more importantly, it contains the typical features of different individuals making different gestures, which makes the method have a stronger individual adaptability. 2. In view of the problem that the traditional gesture interaction method has a large amount of computation in the process of recognition, the clustering method can improve the problem to some extent. At the same time, this paper puts forward the idea of spindle classification. The test gesture sequence can only match the template with the same spindle, which can effectively reduce the time complexity of the online template matching process in the whole gesture interaction system, so as to determine its class. 3. Aiming at the problem that the traditional methods are more vulnerable to the redundant information and noise information in gesture signals resulting in the misjudgment of gesture categories, a compression sensing method is proposed as a means of recognition in gesture interaction. This method can not only reduce the computation, but also restore the gesture signal without distortion, keep the important features, and then improve the recognition rate of the algorithm and obtain the recognition results.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP212

【参考文献】