光流特徵應用於走勢辨認 ; Human Gait Classification using Optical Flow Features

NCU Institutional Repository > 資訊電機學院 > 資訊工程研究所 > 博碩士論文 > Item 987654321/44626

請使用永久網址來引用或連結此文件: http://ir.lib.ncu.edu.tw/handle/987654321/44626

題名:	光流特徵應用於走勢辨認;Human Gait Classification using Optical Flow Features
作者:	鄭建宏;CHIEN-HUNG CHENG
貢獻者:	資訊工程研究所
關鍵詞:	走勢辨認;光流演算法;gait classification;optical flow
日期:	2010-07-19
上傳時間:	2010-12-09 13:50:54 (UTC+8)
出版者:	國立中央大學
摘要:	在圖形識別研究領域中，有許多身份認證的應用，諸如臉部辨識、掌紋辨識、虹膜辨識以及簽名辨識等等，走勢辨識相對於其他辨識系統來說，是比較方便且人性化的，它不需要額外的特殊儀器，所需要的設備僅需要一台攝影機，同時測試者也不必停下來做指定的動作，只需要從架設好的環境經過，即可辨識出他的身份。在近年來監視攝影機普遍的環境下，走勢辨識能夠更有效的去利用監控系統。本篇論文提出以光流當作特徵資訊的系統架構，實做三種不同的光流演算法，並且透過多樣化的測試，來分析與討論其優缺點。我們所提出的系統流程分別如下，首先將輸入的走勢影像轉換成灰階色彩空間，經過光流演算法的計算，取得光流資訊，先大略捨去掉強度較弱的雜訊，使用高斯模型(GM)來逼近光流的分佈，推算出移動物的實際位置，以這個範圍為基準，將光流資訊以強度和角度的表示，建立出二維直方統計圖，並且根據每個人走勢的運動週期來做特徵資訊的正規劃，最後將特徵資訊經過主成份分析(PCA)和線性辨別分析(LDA)後，以k-nn分類器來取得辨識結果。我們利用CASIA走勢資料庫以及自己所錄製的資料庫來驗證我們系統的可行性，其中包含測試系統對於衣著改變以及攝影機距離的影響，經由實驗結果證實，即使每個人的走勢不見得會有顯著的差異，以走勢資訊當作特徵確實可以解決身份認證的問題。此外，即便我們所提出的系統架構僅使用光流資訊作為特徵，其辨識率也能夠和以輪廓為特徵的方法結果不相上下。最後，針對實驗結果我們加以分析及討論，歸納並提出可能的改進方向。Gait classification is an effective and non-intrusive way for human identity identification on the research field of pattern recognition. The required device that we need is a camera, subjects just walk through the hallway naturally and the system can recognize their identity automatically. Recently, surveillance cameras are installed almost everywhere. Gait classification can be used effectively in detecting illegal intruders without installing extra equipments. In this thesis, we adopt optical flow information as the basic features of gait. Three kinds of optical flow algorithms are manipulated on our proposed system and a variety of testing, analyses and discussions are made to highlight the experiment results. We test and verify that the optical flow information is really working not only on moving object detection and tracking but also on the complex problem of gait classification. In our work, we first transform the input color image to gray level space. Then, employ optical flow algorithm to get the optical flow field. Some optical flow information on which intensity values are too small are removed and then Gaussian Model (GM) is employed to model the strong intensity information of optical flow. The locations of foreground objects (gait human) are extracted according to the mean and covariance of GM. After that, we adopt flows to construct feature histograms which belong to the bounding box. Since gait cycles are different with each subject, we apply the histogram matching method to normalize each video feature. Next, Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA) are performed to reduce the feature dimension and find the best projection axis, respectively. Finally, k-Nearest Neighbor (k-NN) classifier is adopted to get the recognition result. Through experimental analysis, we find that gait is a very effective feature to classify human’s identity even though the differences of walking postures between different persons are hard to observe. Although the accuracy of optical flow information is doubtful, we proved that optical flow information is useful for gait classification problem. In this thesis, we adopt optical flow information only and do not consider shape features or other information. Moreover, experimental results demonstrate that the proposed framework contribute as good recognition rates as the contour-based approach in CASIA and our own database.
顯示於類別:	[資訊工程研究所] 博碩士論文

文件中的檔案:

檔案	描述	大小	格式	瀏覽次數
index.html		0Kb	HTML	723	檢視/開啟

在NCUIR中所有的資料項目都受到原著作權保護.

社群 sharing

資料載入中.....