基於偏最小平方法之靜態手勢辨識; Hand Posture Recognition Using Partial Least Squares

NCUIR > College of Electrical Engineering & Computer Science > Graduate Institute of Communication Engineering > Electronic Thesis & Dissertation > Item 987654321/60937

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/60937

Title:	基於偏最小平方法之靜態手勢辨識;Hand Posture Recognition Using Partial Least Squares
Authors:	蔡雨芳;Tsai,Yu-Fang
Contributors:	通訊工程學系
Keywords:	手勢辨識;偏最小平方法;梯度方向直方圖;支持向量機;降維;降低訓練時間;Hand Posture Recognition;Partial Least Squares;Histogram of Oriented Gradients;Support Vector Machine;Dimension Reduction;Training Time Reduction
Date:	2013-07-24
Issue Date:	2013-08-22 12:07:12 (UTC+8)
Publisher:	國立中央大學
Abstract:	手勢辨識之應用可滿足人機互動的需求，故基於影像之手勢辨識為重要的研究議題。以影像為基礎之手勢辨識，易受到環境及人為影響，例如：手勢大小、光線變化及視角變化。此外，手勢辨識系統應降低訓練及測試時間，以因應遇到影像資料量大和特徵維度高之問題。在本論文中提出以偏最小平方法(partial least squares, PLS)降低Histogram of Oriented Gradients (HOG)的維度，再結合支持向量機(Support Vector Machine, SVM)並採用RBF核函式，於訓練階段對降維後的HOG徵描述子進行分類器之訓練。於測試階段，對新進影像擷取HOG特徵描述子，並使用在訓練階段所得的降維係數，進行對測試資料的降維，再將降維後的資料輸入SVM分類器內，以OAA分類法則決定最終辨識結果，此既能減少系統處理時間，又能有效增進辨識率。 The needs of applications of hand posture recognition meets the demand for human-computer interaction, thus the vision based hand posture recognition is an important research topic. However, vision based hand posture recognition is vulnerable to environmental changes or human impacts, i.e. size of hand posture, lighting and view variations. In addition , a hand posture recognition system must process efficiently in training and testing stage to handle the large amount of image samples and high dimensional descriptors. In this paper, we propose to apply partial least squares method (partial least squares, PLS) on dimensions of Histogram of Oriented Gradients (HOG) descriptors, and to train support vector machines (Support Vector Machine, SVM) with RBF kernel function. At the training stage, we perform dimension reduction HOG descriptors and then to train SVM classifiers. In the testing phase, after obtaining of HOG feature descriptors on testing images, and then we use the weight matrix from PLS for dimension reduction. To obtain the final classification result, we use SVM classifiers, followed by OAA decision rule. With these helps , this system can not only reduce both the training and testing time, but also improve the recognition rate.
Appears in Collections:	[Graduate Institute of Communication Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	1370	View/Open

社群 sharing

Loading...