植基於階層式動態貝氏網路之動態手勢辨識法

THUIR > College of Engineering > Department of Industrial Engineering and Enterprise Information > Theses and Dissertations > Item 310901/1386

Please use this identifier to cite or link to this item: http://140.128.103.80:8080/handle/310901/1386

Title:	植基於階層式動態貝氏網路之動態手勢辨識法
Other Titles:	Dynamic Gesture Recognition Based on Hierarchical Dynamic Bayesian Networks
Authors:	董俊良 Tung, Chun-Liang
Contributors:	王偉華 Wang, Wei-Hua 東海大學工業工程與經營資訊學系
Keywords:	階層式動態貝氏網路;動態貝氏網路;隱馬可夫模型;人類行為辨識 Dynamic Bayesian Network;Hidden Markov Model;Human Behavior Recognition;Hierarchical Dynamic Bayesian Network
Date:	2009
Issue Date:	2011-02-25T01:55:05Z (UTC)
Abstract:	最近幾年，在視訊串流中的人類行為辨識已受到電腦視覺研究學者廣泛地研究。人類行為辨識可被視為一種時間變動特徵植(time varying feature data)的分類技術。人類行為辨識的根本問題是在於如何從訓練樣本中學習特定的動作序列，以及發展出訓練和分類的方法來處理每ㄧ類動作樣式類別中細微的特徵值變化。隱馬可夫模型(hidden Markov model)是在電腦視覺領域中處理行為辨識最主要的方法之一。隱馬可夫模型以多樣化數學結構的統計模型來處理順序資料，是一種假設未知參數的馬可夫過程，並成為在狀態空間模型(state-space model)中建立隨機過程及序列模型最重要的方法之ㄧ。然而，隱馬可夫模型僅使用單個離散隨機變數來表示數個觀察資料的狀態分佈，如此將造成高結構複雜度的問題產生。比較而言，動態貝氏網路(dynamic Bayesian network)使用多個隨機變數來表示數個觀察資料的狀態，因此可以被使用來克服高結構複雜度的建模問題。動態貝氏網路為貝氏網路的延展，是一種在特定時期內建立隨機變數間的機率分佈模型的方法。本研究的主要目標為從視訊串流中快速辨識人類行為及動作序列。因此，本研究提出以階層式動態貝氏網路為基礎的階層式動態視覺模型(Hierarchical Dynamic Vision Model , HDVM)架構，做為辨識動作的基礎模型。以色彩為基礎的切割方法(color-based segmentation )將作為追蹤物件的移動軌跡及移動方向的偵測機制，其所產生的資料將提供作為特徵值計算的依據，並將特徵值應用在HDVM模型的參數訓練程序。在模型參數的估算階段中，為避免產生區域最佳參數，本研究亦提出繼承平衡搜尋演算法(Inheritance Balance Searching algorithm, IBS)作為參數訓練的方法。該訓練法將以較佳的全域探勘(global exploration)為導向並使用Forward-Backward 演算法計算特定觀察序列的機率，並且在模型參數的極大化概似度估算中，將以Expectation-Maximization 演算法針對每ㄧ個觀察資料組找尋在區域鑽探(local exploitation)搜尋的最佳模型參數。本研究的實驗結果顯示，以階層式動態貝氏網路為基礎的階層式動態視覺模型可提供較準確及有效率的人類行為及動作辨識。 Human behavior recognition in video stream has been studied extensively from computer vision researchers in recent years. Recognition of behaviors can be thought as the classification of time varying feature data. The fundamental problem of behavior recognition is how to learn the specific motion sequences from training samples, and to develop both training and classifying methods for coping with slight variations of the features data within each class of motion patterns. A hidden Markov model is one of the major existing methods for behavior recognition in the field of the computer vision. The hidden Markov model, a statistical model with rich mathematical structure for ordered data, is assumed to be a Markov process with unknown parameters and which has become the one of the most significant methods for modeling stochastic processes and sequences in the state-space model. However, a hidden Markov model only uses a discrete random variable to represent a number of the distribution of observation states, which may result in a problem of the high structure complexity. By contrast, a dynamic Bayesian network uses a set of random variables to represent a number of observation states, with which can be used to rise above the modeling problem of the high structure complexity. A dynamic Bayesian network, extended of Bayesian network, is a way to model probability distributions among random variables over time. The principal aim of the research is to recognize human behavior and analyze human motion sequence from video stream. Based on the hierarchical dynamic Bayesian network, we propose the Hierarchical Dynamic Vision Model (HDVM) which provides a framework to model and recognize the human motion. It is featured by tracking the moving direction and the trajectory of the object with color-based segmentation which can be applied in the training process of the HDVM. For a better global exploration, the Inheritance Balance Searching algorithm (IBS) is introduced to avoid having the local optimal parameters, with which the Forward-Backward algorithm is used for computing the probability of a particular observation sequence in the parameter estimation. And for the maximum likelihood estimation of model parameters, the Expectation-Maximization algorithm is used to find out the optimal parameters of local exploitation searching for each observation data set. The results of this study show that based on the hierarchical dynamic Bayesian network of the Hierarchical Dynamic Vision Model can provide better recognition accuracy and efficiency for recognizing human behaviors and motion patterns.
Appears in Collections:	[Department of Industrial Engineering and Enterprise Information] Theses and Dissertations

Files in This Item:

File	Size	Format
097THU00030001-001.PDF	957Kb	Adobe PDF	415	View/Open

Loading...