洪一平Hung, Yi-Ping臺灣大學:資訊工程學研究所陳昱廷Chen, Yu-TingYu-TingChen2010-06-092018-07-052010-06-092018-07-052009U0001-1305200914361900http://ntur.lib.ntu.edu.tw//handle/246246/185354視訊安全監控相關研究中，前景偵測與行人偵測是基礎且重要的課題。在本論文裡，我們針對背景模型學習、整體行人偵測和部件行人偵測等題目進行研究與探討。在背景模型學習方面，大多數方法都是以像素為基礎，其優點是可提供精細的前景偵測結果，但對動態背景則無法有好的描述。近年來，有一些研究使用以區塊為基礎的方法來處理動態背景的問題，但無法提供精細的前景偵測結果。有鑑於此，我們提出了一個階層式架構來整合以像素與以區塊為基礎的方法，此架構除可克服動態背景的影響，亦可提供兩階段多尺度的前景偵測結果。在整體行人偵測的研究上，我們使用異質性特徵來描述行人影像，並提出了階層式連鎖前饋分類器來學習整體行人偵測器。在我們的架構中，藉由使用Meta階層，不同階層間的資訊可以被使用，因此，偵測正確率與效率可進一步提昇。當人被遮蔽時，整體行人偵測器可能無法將人成功偵測出來。針對此一問題，我們提出了一個多類多重實例boosting來學習部件偵測器。藉由使用多重實例學習方法，訓練影像對位的問題可以被解決；且藉由特徵共用的特性，我們的方法可以學習出一個有效率的部件偵測器。此外，我們也提出一個機率模型方法來整合部件偵測結果。藉由廣泛的實驗，我們證明此方法能有效地進行人的偵測。最後，我們整合了本論文提出之背景模型技術與行人偵測技術來進行視訊安全監控。Three important research topics in visual surveillance are studied, including background modeling, holistic pedestrian detection, and part-based pedestrian detection. Most previous background modeling approaches are pixel-based, while some approaches began to study block-based representations which are more robust to non-stationary backgrounds. We propose a method that integrates block- and pixel-based approaches into a single framework. Quantitative results show that the proposed method has better classification results than existing single-level approaches. In addition, we develop a method that can detect holistic pedestrians in images. In our approach, heterogeneous features are employed for weak-learner selection, and a novel cascaded structure that exploits both the stage-wise classification information and the inter-stage cross-reference information is proposed. Experiment results show that our approach can detect pedestrians with both efficiency and accuracy. We also propose a multi-class multi-instance boosting method for effective part-based pedestrian detection in images. Training examples are represented as a set of non-aligned instances, and the alignment problem caused by human appearance variation can be handled. Our method has the feature-sharing ability in a cascaded structure for efficient detection. Experiment results demonstrate the superior performance of the proposed method. We also combine background modeling and pedestrian detection techniques for visual surveillance application.口試委員會審定書 iii謝 v要 viibstract ixist of Figures xvist of Tables xix Introduction 1.1 Motivation 1.2 Overview of the Dissertation 3.2.1 Hierarchical Background Modeling and Foreground Detection 3.2.2 Holistic Pedestrian Detection 4.2.3 Part-Based Pedestrian Detection 5.3 Dissertation Organization 6 Hierarchical Background Modeling and Foreground Detection 9.1 Introduction 9.2 Literature Review 10.3 Coarse-Level Modeling 12.3.1 Contrast Histogram of Gray-Level Images 13.3.2 Contrast Histogram of Color Images 15.3.3 Background Modeling by Contrast Histograms 16.3.4 Coarse-Level Experiment Results 18.4 Hierarchical Background Models 19.4.1 A General Description of Pixel-Based Background Modeling 20.4.2 Asymmetric Feed-Forwarding 21.5 Experiment Results 23.5.1 Implementation 23.5.2 Performance Results 23.6 Summary 28 Holistic Pedestrian Detection 31.1 Introduction 31.2 Literature Review 32.3 Real AdaBoost and Feature Pool 35.3.1 Intensity-Based Features 36.3.2 Gradient-Based Features 37.3.3 Combined Feature Pool 40.4 Cascading Feed-Forward Classifiers 42.4.1 Adding Meta-Stages 45.4.2 Meta-Stage Classifier 47.4.3 General Structure with Meta-Stages 47.4.4 Distinction of AdaBoost Stage and Meta-stage 48.5 Experiment Results 51.5.1 Adopted Human Dataset 51.5.2 Implementation 51.5.3 Performance Results of Combining Intensity- and Gradient-Based Features 52.5.4 Performance of Inserting Meta-Stages 58.5.5 Efficiency and Accuracy Comparison with HOG 61.6 Visual Surveillance Application 63.7 Summary 71 Part-Based Pedestrian Detection 73.1 Introduction 73.2 Literature Review 74.3 Real Version MILBoost 77.3.1 A Review of AnyBoost 77.3.2 MILBoost 78.3.3 Real MILBoost 79.4 Multi-Class Multi-Instance Boosting 81.4.1 MCMIBoost: Confidence Value Evaluation 82.4.2 Cascaded MCMIBoost Architecture 83.4.3 Probability Combination Classifier 84.5 Experiment Results 86.5.1 Results on the MIT Dataset 88.5.2 Results on the INRIA Dataset 91.6 Visual Surveillance Application 93.7 Summary 98 Conclusion and Future Work 99.1 Conclusion 99.2 Future Work 101ibliography 103ublications 11113017646 bytesapplication/pdfen-US階層式背景模型差值統計圖人員偵測整體行人偵測階層式前饋分類器meta階層AdaBoost部件行人偵測多重實例學習多類多重實例boosting特徵共用視覺監控Hierarchical background modelingcontrast histogramhuman detectionholistic pedestrian detectioncascaded feed-forward classifiersmeta-stagespart-based pedestrian detectionmulti-instance learningmulti-class multi-instance boostingfeature sharingvisual surveillancethesishttp://ntur.lib.ntu.edu.tw/bitstream/246246/185354/1/ntu-98-D92922013-1.pdf